Universitat Autonoma de Barcelona

Master Thesis

Staff Scheduling in Ground Handling

Company

Student:
Elena Popova

Supervisors:
Dr. Miquel Àngel Piera
Dr. Liana Napalkova

September 2015
Abstract

This master thesis aims to analyze ground handling scheduling task in large scale airline
operations in order to collect results which might be used in contribution to the operational
efficiency of ground handling companies.
This work is focused on designing and developing of evolutionary computation algorithm
in Matlab for solving ground handling staff scheduling problem, which consists in finding the
schedule for all ground handling agents, such that the objective function is minimized and all the
constraints are satisfied. The objective function is built in a way that the number of unserviced
by some reason aircrafts is as small as possible, the workload is distributed equally and situations
when the agent is sent successively to gates located far from each other are mostly avoided. The
constraints that were taken into account are time windows of the tasks, working hours of the
staff and transportation time between the gates.
Since many factors could affect the processes which are performed during the turnaround, it
is not an easy task to make a full simulation, but in this paper was made an attempt to take at
least the main variables into consideration.
The analysis of problem solving methods was done in the theoretical part of the work in
order to choose the one, which is the most appropriate to the specifics of formulated problem.
Finally, two cases corresponding to two different implementations were considered in this the-
sis and their comparative analysis was conducted in order to allow more exhaustive exploration
and validation of the optimization method.

1
Table of contents

1 Introduction 4
1.1 State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.1 Economic Aspects of Ground Handling . . . . . . . . . . . . . . . . . . . . 4
1.1.2 Impact of Turnaround Performance on Flight Delays . . . . . . . . . . . . 5
1.1.3 Manual vs. Automated Scheduling . . . . . . . . . . . . . . . . . . . . . . 6
1.1.4 Commercial Solution (INFORM) . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Structure of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2 Analysis of the Ground Handling Scheduling Problem 9

2.1 Aircraft Ground Handling Activities . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1 Deplaning and Boarding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.2 Baggage Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.3 Cargo Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.4 Aircraft Refueling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.5 Cabin Service . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.6 Catering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.7 Passenger Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.8 Water Supply and Lavatory Drainage . . . . . . . . . . . . . . . . . . . . 12
2.1.9 Supplies of Power, Air-Conditioning and Compressed Air . . . . . . . . . 12
2.1.10 Aircraft De-icing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.11 Pushback Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 The Impact of Uncertainty on Ground Handling Operations . . . . . . . . . . . . 14
2.3 The Planning Horizon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 Airport Collaborative Decision Making (A-CDM) . . . . . . . . . . . . . . . . . . 16
2.4.1 A-CDM Milestones Approach . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5 System Wide Information Management (SWIM) . . . . . . . . . . . . . . . . . . 17

3 Mathematical Formulation of the Problem 20

3.1 Classifying Routing and Scheduling Problems . . . . . . . . . . . . . . . . . . . . 20
3.2 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 Input Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.4 Objective Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.5 Decision Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.6 Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

4 Overview of Problem Solving Methods 24

4.1 Exact methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.2 Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.3 Metaheuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2
TABLE OF CONTENTS

4.3.1 Simulated Annealing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

4.3.2 Tabu Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.3.3 Particle Swarm Optimization . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.3.4 Ant Colony Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.3.5 Evolutionary Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3.5.1 Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3.5.2 Evolutionary Programming . . . . . . . . . . . . . . . . . . . . . 32
4.3.5.3 Evolution Strategies . . . . . . . . . . . . . . . . . . . . . . . . . 33

5 Decision Support Tool 35

5.1 Algorithm Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.2 Initial Population . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.2.1 Case 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.2.2 Case 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.3 Termination Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.4 Tournament Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.5 Genetic Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.5.1 Case 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.5.2 Case 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.6 Fitness Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.6.1 Case 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.6.2 Case 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

6 Case Studies 38
6.1 Input Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
6.2 Case Study 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.2.1 Presentation of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.2.2 Computational Time Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.3 Case Study 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
6.3.1 Presentation of Model Output . . . . . . . . . . . . . . . . . . . . . . . . 46
6.3.2 Computational Time Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 48

7 Analysis of Results 50

8 Conclusions and Future work 52

8.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
8.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

Annex A 56

3
Chapter 1

Introduction

1.1 State of the Art

During the ground time at the airport, an aircraft requires different handling services, e.g.
the unloading and loading of baggage, the cleaning of the aircraft cabin, fuel and water supply,
etc. Previously those operations were performed by a division of the airport.
The Directive 96/67/EC dating from October 1996 [1] has opened up the ground handling
market to competitors to prevent the monopoly of EU airports. This resulted not only in
reducing the costs of ground handling for the airlines but also in improving of the quality of
the services. According to the Directive, in major EU airports, which have more than 2 million
passengers or 50000 tonnes of freight per annum, at least two ground handling suppliers must
be available for certain service categories. At least one of the suppliers must be independent
from the airport and the carrier.
The aircraft turnaround is crucial for airline schedule adherence, for high customer satis-
faction, and economic productivity. Ground staff scheduling is an important area of research,
since even small improvements in staff scheduling could translate into large savings for airline
companies.
This work presents model and algorithms for general optimization and decision problems
arising within ground handling staff scheduling.

1.1.1 Economic Aspects of Ground Handling

Recently, many airports and airlines made the transition in ownership towards privatization,
which caused the airport business environment become more competitive. Airports are in com-
petition to attract airline routes, while airlines are trying to cut their costs. ”We only make
money off our planes when they are in the air” says Chris Wahlenmeier, vice president of ground
operations in Southwest Airlines [2].
The airlines are eager to turn their planes round fast and get them back in the air as soon
as possible. An example could be Ryanairs0 attitude towards hold luggage. Hold luggage takes
time to load and unload; it also adds weight to the plane which results in higher price for fuel
per trip. Less hold luggage means less time for loading/unloading and refueling, therefore less
time to be spend in the airport (not to mention the price of the fuel). In this case Ryanair
charges for hold luggage rather to reduce the turnaround time and therefore ground handling
costs than to get some ”extra” money from customers.
Shorter turnaround time results in more trips per plane with the added benefit of getting
more trips out of airline staff. Therefore ground handling companies are urged to increase cost
effectiveness and to deliver faster and more reliable service. The monetary value of ground
handling services accounts for about 5 to 8 percent of the airline ticket, depending on the type
of airline being used. The global market for ground handling is now estimated to be worth over

4
1.1. STATE OF THE ART

USD80 billion per annum according to its trade association. By comparison, the airline industry
turned over around USD789 billion in 2014.
Since ground handling crews are located at the airport all the time, it is possible to call in
extra hands if needed. This has made the need for a good planning solution less pressing for
ground handling companies than airlines in the past. However, even small improvements in staff
planning of ground handling company could translate into large savings.

1.1.2 Impact of Turnaround Performance on Flight Delays

An airline schedule is rarely implemented as planned. It is often disrupted due to bad weather
conditions, aircraft breakdowns, crew delays, insufficient ground operation performance, etc.
Figure 1.1 shows an average number of departures a day within a month from January 2011
to April 2015 in Europe.

Figure 1.1: Traffic 2011-2015 [3]

Statistics shows that an average delay per delayed flight in April 2015 appeared to be 27
minutes. 40% of delayed flights were delayed on departure and 39% of flights were delayed on
arrival (with a delay time > 5 minutes) [3]. More statistics concerning worldwide flight delays
could be found on www.flightstats.com/go/Media/stats.do/
Reasons for flight delays can be assigned to five main categories, which cover up to 90% of
potential flight delays: Late-arriving Aircraft, Air Carrier, National Aviation System, Extreme
Weather and Security [4].

• Late-arriving Aircraft: the previous flight using the same aircraft arrived late, causing the
present flight to depart late.

• Air Carrier: the cause of the cancellation or delay was due to circumstances within the
airline0 s control (e.g. maintenance or crew problems, aircraft cleaning, baggage loading,
fueling, etc.).

• National Aviation System: delays and cancellations attributed to the national aviation
system, such as heavy traffic volume or air traffic control.

5
1.1. STATE OF THE ART

• Extreme Weather: significant meteorological conditions (actual or forecasted), such as

tornado, blizzard or hurricane.

• Security: delays or cancellations caused by evacuation of a terminal, re-boarding of aircraft

because of security breach, inoperative screening equipment, etc.

On Figure 1.2 is given the distribution of delay causes, according to data of 14 major airlines in
USA.

Figure 1.2: Delay cause by year [4]

As it could be seen from Figure 1.2, on one hand, the ground handling activities (which
are included in category ”Air Carrier Delay”) have a significant influence onto the air transport
performance, i.e. the delay in providing ground handling operations may result in delay of the
flight.
On another hand, the scheduled turnaround process is always disturbed if the airplane does
not arrive at the allocated gate on time. Ground handling agents have to consider the possible
delays to ensure reliable ground processes aiming to not let this delay propagate into the following
flight schedule.
This means that ground handling performance and aircraft performance in terms of delays
are strongly interrelated. That is another reason why ground staff scheduling plays an important
role.

1.1.3 Manual vs. Automated Scheduling

Ground staff scheduling is a very complex task. Planning often involves hundreds of em-
ployees, thousands of work tasks per week and a large number of constraints. Manual planning
is tedious and error-prone task, and it often takes planners several hours or even days to pre-
pare staff plans by hand. Therefore the need for automated scheduling systems is becoming
significant.
Difficulty which appears while designing the automated system is that not all the constraints
can be modeled and not all the information can be put into the optimization model. For exam-

6
1.1. STATE OF THE ART

ple, it is complicated to introduce the information, conveyed by telephone calls, written notices,
conversations between planners and scheduled personnel or by other means of communication.
Most optimization problems are NP-hard, so it is common to omit some restrictions to ob-
tain models that can be solved in reasonable time on the large instances required in real-life
applications.
In some cases it may be impossible to satisfy all the constraints, however computer-based
planning systems could be used to quickly examine different scenarios and find the least unde-
sirable solution. The possibility of fast validation of feasibility can save a lot of time. In most
real-life applications, schedules provided by automated systems are used as a decision support
tool for a planner, but they do not substitute the work of the planner completely.

1.1.4 Commercial Solution (INFORM)

Nowadays most of the airlines, airports and ground handling companies are using automated
ground staff planning systems. In the literature, several scheduling systems related to specific
airlines have been reported, such as Pan Am [Schindler and Semmel, 1993], Air Canada [Nobert
and Roy, 1998] and Alitalia [Felici and Gentile, 2004].
One of the presently used scheduling software which stands out among the others is Ground-
Star (GS), developed by INFORM a company, specializing in intelligent planning and logistics
decision-making software, based on mathematical optimization algorithms. GroundStar is a
suite of software systems for the intelligent support of ground handling operations and airport
processes, which includes advanced optimisation components and powerful graphical user inter-
face. The goal of this software is to build one overall airport management process following the
Airport Collaborative Decision Making concept, which will be discussed in the next chapter.
GS Planning involves planning of aircraft stands, gates, check-in counters, baggage carrousels
and other airport infrastructure elements, thus facilitating the overall integration of these re-
sources in operations with staff and ground support equipment.
GS RealTime is the creation of tasks in accordance with constantly changing flight schedules
and the allocation of airport resources, such as staff, equipment, gates, baggage carrousels, etc.
GS RealTime integrates with multiple sources of information to ensure that full details of the
operation are available. Its optimization strength provides the most efficient usage of resources
and guarantees that contractual commitments are met and provided services are recorded.

Figure 1.3: INFORMs turnaround management system monitors and guides the progress of each aircraft
turnaround in real-time and foresees bottlenecks and possible delays.

7
1.2. OBJECTIVES

GroundStar is currently used by over 50 organizations in more than 150 airports worldwide.
Scheduling software is nowadays required to flexibly adapt to changes in work rules, problem
characteristics and sizes of operations. Clearly, the development of advanced planning systems
is a costly investment. Therefore many airlines, airports and ground handling companies resort
to generic software packages, such as the one presented by INFORM. However, the use of
generic software has a drawback: the solution is often not fully adapted to the realities of that
particular ground handling company using it. On one side, scheduling system could include
”extra” features; on another side, it could lack some. Therefore ground staff scheduling problem
remains actual up to this day.

1.2 Objectives
The objective of this master thesis is to analyze and solve ground handling scheduling prob-
lem in large scale airline operations in order to collect results which might be used in contribution
to the operational efficiency of ground handling companies.
In order to achieve this objective, the following tasks have been formulated:

1. Analyze different aspects of the ground handling staff scheduling problem.

2. Formulate the mathematical model.

3. Investigate problem solving methods.

4. Develop programming code for ground handling staff scheduling problem with the use of
evolutionary algorithms.

5. Perform comparative analysis of results using Design of Experiments (DOE).

1.3 Methodology
In this master thesis ground handling staff scheduling problem was solved with the use
of evolutionary strategy algorithm with local heuristics. Programming code was developed in
Matlab, Release R2012a.

1.4 Structure of the Thesis

The thesis is organized as follows. Chapter 1 is an introductory part which highlights the
importance of ground handling staff scheduling problem and reviews its existing solutions. Chap-
ter 2 provides an overview of the ground handling activities and modern technologies used for
exchanging information between aviation partners. Chapter 3 is devoted to the mathematical
model. Here the optimization problem is presented by an objective function, which should be
minimized given the set of constraints. Chapter 4 reviews the problem solving methods and ex-
plains why in this work evolutionary algorithms were selected. Chapter 5 provides information
about the chosen features of the algorithm. Chapter 6 is dedicated to description of input data
and case studies. In the last chapters 7 and 8 analysis of the results and conclusions are given.

8
Chapter 2

Analysis of the Ground Handling

Scheduling Problem

2.1 Aircraft Ground Handling Activities

Ground handling is a common term to name tasks that should be performed at the aircraft
during its turnaround – the time between its arrival to the gate and its departure to the next
flight. These tasks can be distinguished by ramp operations and passenger services. Tasks on
the ramp include refueling, aircraft cleaning, water supplying, lavatory drainage, baggage and
cargo handling, catering, towing with pushback tractors and deicing. Passenger services mainly
refer to providing of check-in and boarding services, staffing the transfer counters, customer
service counters and airline lounges.
Figure 2.1 shows ground support systems and mobile equipment of Boeing 777 for a typical
turnaround.

Figure 2.1: Boeing 777 being serviced during a turnaround with the help of ground support systems and mobile
equipment [5]

In the next subsections each of GH operations will be reviewed in details in order to give
more thoroughful view on particular tasks that ground handlers are responsible for.

9
2.1. AIRCRAFT GROUND HANDLING ACTIVITIES

2.1.1 Deplaning and Boarding

Disembarking and boarding can be provided by means of:

• stairs carried by the aircraft

• mobile stairs

• passenger bridges.

A combination is also possible, e.g. a bridge attached to the aircraft door behind the crew cabin
and mobile stairs for the rear exit.
There are many advantages of using passenger bridges over the stairs. Firstly, passengers can
leave/board the aircraft more quickly. Secondly, passengers reach their aircraft safely without
getting lost, since their flow path is straight-forward. Thirdly, passengers are protected against
bad weather. At last, while the aircraft is being boarded or deplaned, other servicing activities,
including aircraft refueling and water supply, can be carried out simultaneously. In this case the
movement of the servicing vehicles across the apron is not obstructed. The safety of passengers
is also ensured as contact between them and servicing vehicles is avoided.
However, passenger bridges are more costly than mobile stairs and airline companies are
charged for using them. That is the reason why low cost carriers usually prefer mobile stairs or
even the aircraft0 s own stairs.

2.1.2 Baggage Handling

Baggage handling is becoming a critical activity in ground handling. One of the criteria for
quality of service of the airport is the time passengers have to wait for baggage after disembarking
the aircraft.
At small airports most of the activities connected with baggage handling can be carried out
manually. The baggage is loaded on the baggage cart either directly or from the belt conveyor
and dispatched to the aircraft, where it is loaded in the aircraft cargo holds. At large airports
with the departure of several flights at the same time, sorting of baggage must be provided.
Sorting can be done manually, when the baggage is picked up from the carrousel by the handlers
and loaded on the carts, or automatically. Automated sorting uses baggage tags with bar codes,
which are automatically printed when passengers check-in.
The baggage handling process still has a considerable human content, leading to many health
and safety concerns. Also because of the human factor mishandling can occur, leading to a
certain number of lost or mislaid bags, which damages the reputation of the airport.

2.1.3 Cargo Handling

Most of the freight is transported through only few large hubs: Memphis, Hong Kong, Tokyo,
Frankfurt, etc. Even though there is an impressive range of available freighters, more than 60
per cent of world freight is taken on passenger aircrafts.
Cargo handling includes movement of goods from landside to airside and vice-versa, or from
aircraft to aircraft. There are 3 levels of mechanization to be provided for cargo handling:

• manual: manpower plus fork lift trucks

• semi-mechanized: roller beds or conveyors

• fully mechanized: Elevating Transfer Vehicles (ETV), Automatic Storage and Retrieval
Systems, Transfer Vehicles.

10
2.1. AIRCRAFT GROUND HANDLING ACTIVITIES

The fully mechanized approach only really works with high volumes of containerized freight
and in a setting that can guarantee good maintenance skills. Even so, the whole terminal can
come to a halt if an ETV breaks down. Transfer Vehicles are powered roller conveyors mounted
on a carriage with bogies which are electrically driven along rails in response to a command
system. Mechanization is much more expensive in terms of first cost, but reduces the labor
requirement, causes less handling damage and less risk of mishandling.

2.1.4 Aircraft Refueling

Distribution of fuel from the fuel storage at the airports into the airplanes may be exe-
cuted in two ways: by tankers or by hydrants. Both of the systems have their advantages and
disadvantages.
The principal advantage of the fixed fuel delivering system is a reduction of operating costs
and an increase of safety during aircraft refueling. There are no tankers on the apron with a large
quantity of highly inflammable fuel, but only a dispenser with fuel in hoses from the hydrant
to the airplane and in the filters. The fuel supply to a single Boeing 747 usually amounts to
100 m3 or 80 tons of fuel. With a fixed fuel delivering system, in the majority of cases, only
one worker is needed to perform refueling. In comparison, for the same operation, two or three
tankers with operators are required.
The investment costs of constructing the hydrant system are several times higher than those
of a purchase of several tankers. Nevertheless, its operating costs are lower and its service life is
longer. Even though a hydrant system of fuel distribution has been constructed in the airport,
it is necessary to have available several tankers in case of failure or maintenance of a part of the
hydrant system, so that fuel can still be supplied.

2.1.5 Cabin Service

Cabin services ensure passenger comfort. There are basically three levels of cleaning services
of the aircraft: quick transit cleaning, overnight cleaning and deep cleaning.
The types of services depend on the duration of turnaround. The tasks in a typical quick
transit cleaning involve: (1) seat cleaning, (2) seat pocket cleaning, (3) galley cleaning, (4) toilet
cleaning and replenishment, (5) floor cleaning and (6) blanket management. Overnight cleaning
is more thorough and in addition to the above processes includes: (1) floor vacuuming, (2)
window cleaning, (3) stowage cleaning and (4) cleaning of the cabin crew resting area. Deep
cleaning includes all the above tasks but in a more extensive manner [6].

2.1.6 Catering
Catering includes unloading of unused food and drinks from an aircraft, and loading of
fresh food and drinks for passengers and aircraft crew. Airlines meals are typically delivered in
trolleys. Empty or trash-filled trolleys from the previous flight are replaced by fresh ones. Meals
are usually prepared on the ground in order to minimize the amount of preparation (apart from
chilling or reheating) required in the air.
While some airlines provide their own catering, others have outsourced their catering to
third-party companies.

2.1.7 Passenger Services

Passenger handling includes following services:

• Providing check-in counter services for the departing passengers, i.e. accepting baggage
that has to go in the aircraft0 s cargo hold and issuing boarding passes. Some low cost
airlines tend to use online check-in, which is the process when the passengers confirm

11
2.1. AIRCRAFT GROUND HANDLING ACTIVITIES

their presence on a flight via Internet and typically print their own boarding passes. By
Ryanairs0 policy, passengers who have lost/forgotten their printed boarding passes are
required to pay a Boarding Card Reissue Fee in order to receive a new one at the airport.

• Providing gate arrival and departure services. The handling agents are required to meet
a flight on arrival and guide passengers to baggage claim areas and terminal exit, as well
as provide departure services including boarding passengers, closing the flight, etc.

• Assistance of arriving and departing VIP, unaccompanied children, disabled and elderly
passengers from the point of arrival at the terminal through the whole process to the
aircraft boarding

• Staffing the Transfer Counters, Customer Service Counters, Airline Lounges, Lost and
Found, etc.

2.1.8 Water Supply and Lavatory Drainage

Aircraft are supplied with potable and non-potable water for lavatory sink use through
temporary connections at the airports. Ground handlers typically board aircraft water using
designated watering points (facilities where water is transferred from a water supply to an
aircraft, including water cabinets, carts, trucks and hoses).
Each aircraft equipped with a bathroom or lavatory needs to expend its waste somehow.
After the aircraft arrives it is the lavatory agent’s job to flush the lavatory system.

2.1.9 Supplies of Power, Air-Conditioning and Compressed Air

Most aircrafts can meet their energy requirements on a ramp with a use of an Auxiliary
Power Unit (APU). Advantages of such a solution are the independence of the ground source
and savings on time required to connect/disconnect the ground source. However, the total costs
of APU operations, including maintenance costs, are often higher than the charges for using
external ground sources. Other disadvantages are low efficiency of the APU (only about 30%),
and high noise and environmental pollution due to exhaust gases. At Copenhagen airport, APUs
have to be turned off within 5 minutes after the plane docks and at Zürich airport this time
period is reduced to 30 seconds. After that, the aircraft has to be supplied with an external
electrical power.
If there is no ground support system supplying air-conditioning, the pilot cannot avoid using
APU in order to provide ventilating, heating or cooling in the cabin. Air-conditioning has to
be provided even in mild latitudes. Due to the high concentration of people aboard, cooling of
the cabin has to be assured when the temperature rises above 10 ◦ C. When the temperature is
between 4.4–10 ◦ C air has to be circulated, and when it drops below 4.4 ◦ C the cabin has to be
heated. Central distribution systems can supply air of -6.5 ◦ C in summer and +66 ◦ C in winter.
In order to start the engines compressed air of 230 C◦ is supplied [7].

2.1.10 Aircraft De-icing

It is dangerous for an aircraft to take off while it is covered by snow or ice. Creating sufficient
lift requires high values of the lift coefficient. Research has shown that even a 0.5 mm layer of
ice covering the whole upper area of the wing can decrease the maximum lift coefficient by up
to 33%. Because of this in the majority of countries it is strictly required that de-icing is carried
out prior to take-off.
According to the Association of European Airlines recommendations [8] there are two differ-
ent procedures for the aircraft surfaces treatment:

12
2.1. AIRCRAFT GROUND HANDLING ACTIVITIES

• De-icing is a procedure by which frost, ice, slush or snow is removed from an aircraft in
order to provide clean surfaces

• Anti-icing is a precautionary procedure which provides protection against the formation

of frost or ice and accumulation of snow or slush on treated surfaces of the aircraft for a
limited period of time.

For the de-icing procedures, fluids based on glycol are mostly used. However, if an aircraft
is covered with thick layers of snow, slush, etc. some airline companies are using an effective
and cheap ”mechanical” method – snow is simply brushed away.

2.1.11 Pushback Operations

Pushback is a procedure during which an aircraft is pushed backwards away from an airport
gate by external power. Pushbacks are carried out by special, low-profile vehicles called pushback
tractors or tugs.
Although many aircraft are capable of moving themselves backwards on the ground using
reverse thrust, the resulting jet blast may cause damage to the terminal buildings or equipment.
Engines close to the ground may also blow sand and then suck it into the engine, causing damage
to it. A pushback is therefore the preferred method to move the aircraft away from the gate.
Figure 2.2 shows a typical turnaround of a commercial aircraft at the airport. As we can see,
many tasks are performed simultaneously. The considered turnaround time is 60 minutes, but
this time can be much shorter in the case of regional jets. Usually, ground handling operations
are supposed to be finished within 20 minutes which leaves no room for errors.

Figure 2.2: Example of ground activities for a commercial aircraft [9]

13
2.2. THE IMPACT OF UNCERTAINTY ON GROUND HANDLING OPERATIONS

2.2 The Impact of Uncertainty on Ground Handling Operations

The tighter the operational schedule is, the more important is the operational management.
Figure 2.3 shows that sometimes ground handlers have to deal with uncertainty. If a breakdown
occurs because of an equipment failure, the operation manager has to make a choice depending
on the importance of the interrupted operation. If the flight does not rely on this operation,
the manager may decide to cancel or to shorten it. For example in case of freight/mail loading,
delays of the aircraft may cost more money than loading can make. The problem considered on
Figure 2.3 is the breakdown of the cargo loader.

Figure 2.3: Effect of breakdown and delay on apron dispatch [10]

Figure 2.3 (a) represents normal activity, when no control action is required. Figure 2.3
(b) represents a delay appeared through a breakdown. In this case control is required and is
performed through the following actions:
• Assess the nature of the problem and realize how much time the problem will take to be
sorted out.
• Take corrective action: call equipment base and ask engineer to come to the aircraft or
call up a replacement loader.
• Advise all other activities which will be affected by the breakdown and give them instruc-
tions as necessary (e.g., notify movement control of a delay, tell passenger service to delay
boarding, etc.).

14
2.3. THE PLANNING HORIZON

2.3 The Planning Horizon

Ground handling staff scheduling gives rise to a number of challenging optimization problems.
The workload is naturally represented by a list of work tasks, which should be performed by
employees of appropriate qualification. It is typical to classify staff scheduling problem according
to temporal dependencies into strategic, tactical, operational and real-time planning (Figure 2.4).
These stages are strongly interrelated and different optimization problems have to be solved at
each of them.
Strategic planning, also known as demand planning, is the most distant from the day of
operation and is mostly associated with long-term decision making. It could significantly change
the size and composition of the workforce by hiring or firing personnel, equipment acquisitions
and scenario analysis for bidding on new contracts.
Tactical planning is concerned with mid-term problems that determine availability on the
day of operation from a pool of resources that has been predetermined during the strategic
planning step. During tactical planning shift duties for future time period, such as next month
or next season, are generated. Therefore this phase is sometimes referred to shift planning.
Operational planning deal with an actual day of operation and it mainly covers daily rostering
problems in task scheduling. These could be calculated at the beginning of the day or few days
before the day of operation. At this phase, the demand and the resource availability are fixed.
The problem is then to assign the work to available personnel as efficiently as possible under the
set of constraints specified by the previous planning steps. Some of staff management decisions,
such as shift swaps, sickness or overtime handling, may be viewed as operational.
Real-time planning, also known as dynamic planning, is the final stage of staff scheduling.
It is associated with adapting an existing plan for the day of operation to handle failures that
may occur during the day, such as flight delays or equipment breakdowns.

Figure 2.4: Planning timeline

In this thesis operational ground staff scheduling will be considered, the estimation of the
workload and the generation of staff timetables will be done at the beginning of the day or
several days before the day of operation. This means that the workload for the planning period
is calculated from a fixed flight schedule and known Service Level Agreements (SLAs) between
airlines and ground handling companies. SLA set the level of quality of the handling services,
including strictly definition of their order and time windows. In the event of failure to meet
specified standards, ground handling companies are charged compensation.

15
2.4. AIRPORT COLLABORATIVE DECISION MAKING (A-CDM)

As for example, the performance standards for delivery times of incoming luggage at Manch-
ester Airport, agreed between airlines and ground handling agents are the following:

Aircraft seat numbers Maximum delivery times (in minutes)

First bag Last bag

180 or more 25 40
85 to 179 20 30
40 to 84 15 25
Up to 40 12 15

where the time refers to minutes after ATA (Actual Time of Arrival).
Source: www.caa.co.uk/docs/5/ergdocs/ccreportma/chapter5.pdf.
Real-time aspects like short-term schedule changes, flight delays or personnel sickness could
also be taken into account by rerunning the algorithm with new input data or by introducing
manual changes in the schedule.

2.4 Airport Collaborative Decision Making (A-CDM)

The activities of airlines and airports are complementary in nature but the industry is in need
of better coordination between all the aviation partners. Airport Collaborative Decision Making
(A-CDM) is a concept aiming to improve the operational efficiency of all airport operators
by reducing delays, increasing the predictability of events during the process of a flight and
optimizing the utilization of resources. This aim is to be achieved via improved real time
information sharing between airport operators, ground handlers, airlines and air traffic control
and via adopting coordinated operational procedures, automatic processes and user friendly
tools.
Figure 2.5 illustrates the key objectives of each of the partners involved in the Collaborative
Decision Making.

Figure 2.5: Partners involved in A-CDM and their objectives [11] ATC – Air Traffic Control, CFMU – Central
Flow Management Unit

Airport CDM is now implemented in 35 European airports, the list could be found on
www.euro-cdm.org/airports.php.
As an example, for Paris Charles de Gaulle airport, where A-CDM has been implemented
since 2010, in 2013 departure times were respected more than 85% of the cases, against 80%

16
2.5. SYSTEM WIDE INFORMATION MANAGEMENT (SWIM)

before, the ground traffic became more fluid (taxiing times have been shortened by 2 to 4
minutes), daily consumption of fuel reduced in 14.5 tons [12].
Information sharing is the most powerful concept element in A-CDM. Clearly, there is a need
of proper managing of all that information. This could be done by System Wide Information
Management or SWIM – an enabler for A-CDM, which will be discussed in one of the next
sections.

2.4.1 A-CDM Milestones Approach

One of the essential procedure characteristics defined by A-CDM is the utilization of so
called ”Milestone Approach”, which involves breaking the global process sequence down into
milestones.
Milestones are used to mark specific points along turnaround timeline, starting from the
planning of the inbound flight from the outstation until the take-off of the flight at the subject
airport. The achievement or not of the individual milestones is shared and distributed among
CDM partners, allowing them to identify possible deviations from schedule and appropriately
respond to them. When milestone events do not appear as planned, prompting mechanism raises
alarm, which triggers the reaction of the responsible partner to provide the re-planning of the
operation. To give an example of how the Milestones Approach may work, imagine a situation
in which boarding has not started 20 minutes before estimated departure time. In this case, the
milestone event indicating ”start boarding” does not appear as planned, triggering an alarm to
inform the involved partners about the missed target. Subsequently, Ground Handlers may be
prompted to confirm the expected duration of the delay, resulting in an updated set of milestones
for the remainder of the process. As a result, all partners will be informed immediately about
the late boarding, enabling them to appropriately respond to the current situation [13].

2.5 System Wide Information Management (SWIM)

Up to now, the management of information used by different members of the Aviation
Community has been evolving independently. As a result of this approach, today0 s Air Traffic
Management (ATM) information systems are insufficiently integrated, resulting in difficulties in
on-time use of information.
In 1997 Eurocontrol presented the System Wide Information Management (SWIM) concept
to the Federal Aviation Administration, where it has been under development ever since. SWIM
is the infrastructure that will allow facilitating greater sharing of ATM system information, such
as airport operational status, weather information, flight data, status of special use airspace and
National Airspace System restrictions.
System Wide Information Management is essential for providing the most efficient use of
airspace, managing air traffic around weather, and, what is crucial for ground handling compa-
nies, increasing common situational awareness on the ground.
Operational stakeholders do not only need to exchange network operations information but
they also need computer-to-computer interface. The Network Manager (NM) aims at providing
consolidated interactive interfaces enabling more effective and collaborative decision making
between all stakeholders. In other words, architecture of the NM Services is provided to support
the interoperability requirements of the SWIM network. In order to address different customers
and needs, the Network Manager0 s Operations services and data are available via standardized
system-to-system interfaces through the NM B2B Web Services. Since it is entirely based on
open standard web technologies, the installation of proprietary software on the customers side
is not required. The customers targeted are mainly Air Navigation Service Providers (ANSPs),
Aircraft Operators (AOs), Airports, Handling Agents and Computerized Flight Plan Service
Providers (CFSPs) (Figure 2.8).

17
2.5. SYSTEM WIDE INFORMATION MANAGEMENT (SWIM)

Figure 2.6: Sharing information today [14]

Figure 2.7: Sharing information tomorrow [14]

The NM B2B Web Services are provided in two ”flavors”: SOAP (Simple Object Access
Protocol) web services and non-SOAP web services. The payload is always XML (Extensible
Markup Language). Airspace data is predicated on the use of open data exchange standards.
Within the aviation community, these are primarily Aeronautical Information Exchange Model
(AIXM), Weather Information Exchange Model (WXXM) and Flight Information Exchange
Model (FIXM). All services make use of HTTPS (Hypertext Transfer Protocol). Access to NM
B2B Web Services requires strong authentication with digital certificates.
Up to end of 2014, 180 organizations have been entitled to access the NM B2B Web Services.
More than one hundred from which access the NM B2B Web Services daily [15].

To sum up, Airport Collaborative Decision Making (A-CDM) and System Wide Information

18
2.5. SYSTEM WIDE INFORMATION MANAGEMENT (SWIM)

Figure 2.8: NM B2B Web Services, customer segmentation [15]

Management (SWIM) are the two modern technologies for managing and real time sharing
information between aviation partners in order to improve the operational efficiency of each of
them. A-CDM and SWIM are not yet implemented in every airport, but from year to year they
are improved and spread up through Aviation Community.

19
Chapter 3

Mathematical Formulation of the

Problem

3.1 Classifying Routing and Scheduling Problems

The classification of routing and scheduling problems depends on certain characteristics of
the service delivery system. The simplest case is Traveling Salesman Problem (TSP), which
asks the following question: Given a list of cities and the distances between each pair of cities,
what is the shortest possible route that visits each city exactly once and returns to the origin
city? The nodes may be visited in any order, the travel costs between two nodes are the same
regardless of the direction traveled, there are no delivery-time restrictions and vehicle capacity
is not considered. An extension of TSP, referred to as the Multiple Traveling Salesman Problem
(MTSP), occurs when there are several vehicles to be routed from a single depot. If the capacity
of multiple vehicles is restricted and there exists a possibility of having varying demands for each
customer, the problem is classified as a Vehicle Routing Problem (VRP). If the customers being
serviced have time restrictions, then the problem is referred to as Vehicle Routing Problem with
Time Windows (VRPTW).
In this thesis the ground handling staff scheduling problem is presented as a Vehicle Routing
Problem with Time Windows, since the aircrafts have to be serviced during time windows
specified by Service Level Agreement. To simplify the model and the solution process, the
complete problem is separated into multiple VRPTWs, which are identified for each worker
qualification and are solved individually. Later on mathematical model will be considered for
single VRPTW for specific qualification.

Figure 3.1

20
3.2. NOTATION

3.2 Notation

NW number of workers
W = {w1 ...wN W } set of workers
NT number of tasks
N Ti number of assigned tasks to worker i ∈ W
T = {t1 ...tN T } set of tasks
Ti = {ti1 ...tiN Ti } ⊆ T scheduled set of tasks performed by worker i ∈ W
NG number of gates
{0} depot
G = {g1 ...gN G} set of gates
Gi = {g1i ...gN i
Ti } ⊆ G scheduled set of gates visited by worker i ∈ W
aij time needed to move from gate i ∈ G to gate j ∈ G
bi duration of shift of worker i ∈ W
dij duration of task j ∈ Ti , i ∈ W
[sij , fji ] scheduled time of the task j ∈ Ti performed by worker i ∈ W
[csij , cfji ] initial time window for task j ∈ Ti , i ∈ W
[bsi , bf i ] shift hours of worker i ∈ W

3.3 Input Parameters

In the proposed model it is necessary to have information from Service Level Agreements
signed between airlines and ground handling companies and information about flights, workers
and gates, which is the following:

• Flights
Timetable, which includes name of the airline, type of the aircraft, numbers of arrival and
departure flights, Scheduled Time of Arrival (STA), Scheduled Time of Departure (STD)
and the number of gate, where the aircraft will be serviced during its turnaround.

• Workers
Identification number of worker, his qualification, tasks that he is qualified to perform and
his working hours.

• Gates
Type of the gate (finger or remote) and the time needed to move from any gate to any
other.

• Service Level Agreement (SLA)

SLA defines strictly the order, time and duration of handling services delivered to the
aircrafts.

21
3.4. OBJECTIVE FUNCTION

3.4 Objective Function

In this thesis, the multi-objective approach for ground staff scheduling is introduced. Here-
inafter, multi-objective approach refers to the weighted sum of performance indicators, not the
Pareto based approach.
The aim is to create an optimal schedule for the workers that service aircraft at an airport
during one working day. To attain that we need to minimize the following function:

F itness = α1 ∗ F1 + α2 ∗ F2 + α3 ∗ F3 + α4 ∗ F4 , (3.1)
where α1 − α4 are weighting coefficients and F1 − F4 are objective functions.
 
X
F1 = max tti = max a0gi + agi gi + agi 0  (3.2)
i∈W i∈W 1 j j+1 N Ti
j=1..N Ti −1

tti – total traveling time of worker i ∈ W – is the time that worker spend on moving from
gate to gate during the day. By minimizing maximum of it, we are avoiding sending the same
agent successively to gates located far from each other.
 
X
F2 = max wti = max bi − dij − tti  (3.3)
i∈W i∈W
j=1..N Ti

wti – total waiting time of worker i ∈ W – is the time that worker spend between performing
the tasks during the day. By minimizing maximum of it, we achieve equally distribution of the
workload between the workers of the same qualifications.
XX
F3 = P enalty1 = 1, f or i, j | fji < cfji (3.4)
i∈W j∈Ti
XX
F4 = P enalty2 = 1, f or i, j | bfi < cfji (3.5)
i∈W j∈Ti

P enalty1 – is the number of planes scheduled to be serviced out of required time windows.
P enalty2 – is the number of planes scheduled to be serviced out of workers0 shift.
(P enalty1 + P enalty2) is the total number of unserviced aircrafts, hence by minimizing it
we are leaving as few tasks uncompleted as possible.

With the aim to achieve a more flexible and accurate decision process, we are looking for a
solution representing the best compromise between the four objectives listed above.

3.5 Decision Variables

Scheduled set of tasks performed by workers Ti = {ti1 ...tiN Ti } ∈ T ∀i ∈ W are the decision
variables in our mathematical model.

22
3.6. CONSTRAINTS

3.6 Constraints
The following constraints should be respected:

• working hours of the staff

bsi ≤ sij ≤ bf i ∀j ∈ Ti ∀i ∈ W (3.6)

bsi ≤ sij + dij ≤ bf i ∀j ∈ Ti ∀i ∈ W (3.7)

• time windows of the tasks

csij ≤ sij ≤ cfji ∀j ∈ Ti ∀i ∈ W (3.8)

csij ≤ sij + dij ≤ cfji ∀j ∈ Ti ∀i ∈ W (3.9)

Time windows information is taken from Service Level Agreement (SLA).

• transportation time between the gates

bsi + a0gi ≤ si1 ∀i ∈ W (3.10)

1

fji + agi gi ≤ sij+1 ∀j = 1..(N Ti − 1) ∈ Ti ∀i ∈ W (3.11)

j j+1

fji ≤ bfi − agi ∀i ∈ W (3.12)

N Ti 0

where
fji = sij + dij ∀j ∈ Ti ∀i ∈ W (3.13)

Furthermore, some tasks may require several workers to cooperate. Cooperating agents have
to initiate work on the task simultaneously, which also has to be maintained in the schedule.

23
Chapter 4

Overview of Problem Solving

Methods

Combinatorial optimization problems (COP) can be regarded as a search for the best element
in some set of discrete items. Determining the optimal solution of Vehicle Routing Problem with
Time Windows (VRPTW) is a NP-hard problem in combinatorial optimization, hence there is
no algorithm that can solve all problem instances efficiently in polynomial time. In principle,
exact algorithms can be used for solving relaxed problems, with smaller amount of input or with
less number of constraints. But as the size of NP-hard problem increases, the time needed to
solve it grows very fast. Therefore commercial solvers tend to use heuristics to tackle real world
VRPTW due to its size and the frequency that it may have to be solved.

4.1 Exact methods

Exact algorithms can range from very general techniques such as tree search with branch-
and-bound to highly problem-dependent algorithms such as Dijkstra0 s algorithm for computing
shortest path between two nodes in a weighted graph.
The exact methods for the VRPTW can be classified into three categories: Lagrange relaxation-
based methods, column generation and dynamic programming.
There are several papers using slightly different approaches to Lagrange relaxation-based
methods. For example, there is variable splitting approach followed by Lagrange relaxation
[Fisher et al., 1997], K-tree approach followed by Lagrange relaxation [Holland, 1975], and side
constraints approach followed by Lagrange relaxation [Kohl and Madsen, 1997].
The column generation approach for solving the VRPTW was used for the first time in
[Desrosiers et al., 1984], and more effective version of the same model used to solve more instances
to optimality was presented in [Desrosiers et al., 1992].
The dynamic programming approach for VRPTW was mentioned for the first time in [Kolen
et al., 1987]. The algorithm presented in [Kohl and Madsen, 1997] uses the branch-and-bound
method to achieve optimality. Branch-and-bound is based on an intelligent enumeration of the
”candidate” solutions to an optimal by successively partitioning the solution space and cutting
the search tree while considering limits that are calculated during the enumeration. The detailed
overview of these techniques is given in [16].
We would not need any other kind of algorithm if there was not the problem that exact
algorithms can take exponential time to terminate.

24
4.2. HEURISTICS

4.2 Heuristics
A heuristic is defined by [Reeves, 1995] as a technique which seeks good (near-optimal)
solutions at a reasonable computational cost without being able to guarantee optimality, to
state how close to optimality a particular feasible solution is or, in some cases, even to guarantee
feasibility.
Heuristic algorithms that set routes from scratch are denoted as route-building, while al-
gorithms that try to improve solution on the basis of already built solutions are called route-
improving.
The first paper on route-building heuristics for the VRPTW is [Baker and Schaffer, 1989].
Their algorithm is an extension of the saving heuristic of the VRP [Clarke and Wright, 1964].
A similar heuristic based on the savings algorithm was developed in [Solomon, 1987] and [Lan-
deghem, 1988]. Another heuristic described in [Solomon, 1987] is a time-oriented, nearest-
neighborhood heuristic.
A problem of building one route at a time is usually that the routes generated in the latter
part of the process are of poor quality as the last unvisited customers tends to be scattered over
the geographic area. [Potvin and Rousseau, 1993] try to overcome this problem of the insertion
heuristics by building several routes simultaneously. The initialization of the routes is done
by using the insertion heuristic of [Solomon, 1987]. This method is better than the Solomon
heuristics but still the solutions are quite far away from optimum. [Russell, 1995] and [Antes
and Derigs, 1995] elaborate further on the insertion approach. In general, building several routes
at the same time results in better solutions than building the routes one by one [17].
In general the heuristics, as f.e. the ones of [Solomon, 1987] and [Landeghem, 1988], return
a solution fast. However, their solutions usually lack in quality: most of the time they differ
from optimum for more than 10%.

4.3 Metaheuristics
Metaheuristics is an iterative strategy designed to find, generate, or select a heuristic that
may provide a sufficiently good solution to an optimization problem, especially with incomplete
or imperfect information or limited computation capacity [18].
Metaheuristics do not guarantee that a globally optimal solution can be found on some
class of problems. However, by searching over a large set of feasible solutions, metaheuristics
can often find good solutions with less computational effort than basic heuristics and other
methods. Therefore, the success of these methods is the ability to ”solve in practice” some hard
combinatorial problems.
There are a wide variety of metaheuristics and a number of properties along which one can
characterize them.
One approach is to characterize the type of search strategy. One type of search strategy
is an improvement on simple local search algorithms: Simulated Annealing, Tabu search, It-
erated Local Search, Variable Neighborhood Search and Greedy Randomized Adaptive Search
Procedure (GRASP). The other type of search strategy has a learning component: Ant Colony
Optimization, Evolutionary Computation and Genetic Algorithms.
Metaheuristics could also be classified by single solution vs. population-based searches. Sin-
gle solution approaches focus on modifying and improving a single candidate solution: Simulated
Annealing, Iterated Local Search, Variable Neighborhood Search and Guided Local Search.
Population-based approaches maintain and improve multiple candidate solutions, often using
population characteristics to guide the search: Evolutionary Strategies, Genetic Algorithms and
Particle Swarm Optimization.

25
4.3. METAHEURISTICS

4.3.1 Simulated Annealing

Simulated Annealing (SA) is the oldest metaheuristic that implemented an explicit strategy
to avoid local minima. The method has its roots in statistical mechanics (Metropolis algorithm)
and it was first presented as a search algorithm for combinatorial optimization problem in [19]
and [20].
Simulated Annealing is inspired by the process of annealing in metallurgy, where a material
is heated into a liquid state and then cooled back into a recrystallized solid state. When using
Simulated Annealing, instead of searching for the best solution in the neighborhood, one simply
takes solution from the neighborhood randomly. In case it is better, it is always accepted as a
new current solution, but if the solution is worse - it is only accepted with a certain probability,
which is determined by the gradually decreasing temperature.
The commonly used formula for acceptance probability is:

e − e0
 
P (accept) = exp , (4.1)
T
where T is the current temperature, e is the energy (or cost) of the current solution and
e0 is the energy of a candidate solution being considered. By reducing the temperature, the
acceptation of new solution becomes more and more selective.
The size of the neighborhood considered in generating candidate solutions can change over
the time or be influenced by the temperature, starting initially broad and narrowing while the
algorithm is executed.
Figure 4.1 provides a pseudo code listing of the main Simulated Annealing algorithm.

Figure 4.1: Pseudo code for Simulated Annealing [21]

The convergence proof suggests that the system will always converge to the global optimum if
cooling period is long enough. The downside of this theoretical finding is that for some problems
the number of samples taken for optimum convergence to occur can be more than a complete
enumeration of the search space [21].
Restarting the cooling schedule with the best solution found so far can lead to improvements
in performance for some problems. The effectiveness of the algorithm and its computation
time could be improved by using a problem specific heuristic in choosing starting point for the
search, by parallelization or by hybridization with other algorithms such as Tabu Search or
Genetic Algorithms.

26
4.3. METAHEURISTICS

4.3.2 Tabu Search

Tabu Search basic ideas were first introduced in [22], based on earlier ideas formulated in
[23]. Tabu Search explicitly uses the history of the search, both to escape from local minima
and to implement an explorative strategy.
At each iteration, the neighborhood of the current solution is explored and the best solution
there is selected as the new current solution. The current solution is set to the best solution
in the neighborhood even if it is worse. This has to be done in order to allow the algorithm to
escape from a local optimum.
To prevent cycling, it is forbidden to choose solutions that were visited recently. Latter are
added to Tabu list. Often, Tabu list does not contain illegal solutions, but forbidden moves. It
makes sense to allow the Tabu list to be overruled if this leads to an improvement of the current
overall best solution. Criteria used for overruling the Tabu list are called aspiration criteria.
Tabu Search is usually stopped after a constant number of iterations without any improvement
of the overall best solution or after a constant total number of iterations.
Candidates for neighboring moves can be generated deterministically for the entire neigh-
borhood or the neighborhood can be stochastically sampled to a fixed size, trading off efficiency
for accuracy.
Figure 4.2 provides a pseudo code listing of the main Tabu Search algorithm.

Figure 4.2: Pseudo code for Tabu Search [21]

In spite of the success that Tabu Search has, there is not yet available theoretical support.
No study made it possible to prove the convergence of the method. In order to improve the
procedure, techniques of parallelization and combination with other approaches are proposed.
The hybridization of Tabu search and Simulated Annealing was proposed in [Osman, 1993] and
[Tuyttens et al., 1994], and the hybridization of Tabu Search and Genetic Algorithm in [Glover
et al., 1995].

27
4.3. METAHEURISTICS

4.3.3 Particle Swarm Optimization

The Particle Swarm Optimization (PSO) is a recently proposed metaheuristic method [Kennedy
and Eberhart, 1995], [Eberhart et al., 2001] that simulates the collective behavior of the animal
displacements, e.g. the fish school or bird flock, where the individuals themselves have access
only to limited information, such as position and the speed of their closest neighbors. It can be
observed that a fish school is able to avoid a predator in the following manner: initially it gets
divided into two groups, then the original school is reformed (Figure 4.3), while maintaining the
cohesion among the school.

Figure 4.3: A schematic of a fish school avoiding predator. (a) the school forms only one group, (b) the
individuals avoid the predator by forming a ”fountain” like structure, (c) the school is reformed [24]

The standard PSO algorithm works by having a population of candidate solutions, which
move around in the search space according to the particle0 s position and velocity. The movement
of a particle is guided by its own best-known position and the entire swarm0 s best-known position
in order to move all the particles toward the global best solution in the solution space. At each
iteration, the swarm0 s best-known position is updated. The process is repeated for a given
number of iterations.
Each iteration a particle0 s velocity is updated using:
 
vi (t + 1) = vi (t) + c1 ∗ rand() ∗ (pbest
i − pi (t)) + (c2 ∗ rand() ∗ (pgbest − pi (t))) , (4.2)

where vi (t + 1) is the new velocity for the ith particle, c1 and c2 are the weighting coefficients
for the personal best and global best positions respectively, pi (t) is the ith particle0 s position at
time t, pbest
i is the ith particle0 s best known position, and pgbest is the best position known to
the swarm. The rand() function generates an uniformly random variable ∈ [0, 1].
A particle0 s position is updated using:

pi (t + 1) = pi (t) + vi (t). (4.3)

Figure 4.4 provides a pseudo code listing of the standard Particle Swarm Optimization
algorithm.

28
4.3. METAHEURISTICS

Figure 4.4: Pseudo code for Particle Swarm Optimization [21]

The number of particles should be low, around 20-40. The speed with which a particle can
move (maximum change in its position per iteration) should be bounded, such as to a percentage
of the size of the domain. Particles may leave the boundary of the problem space and may be
penalized, be reflected back into the domain or biased to return back toward a position in the
problem domain. An inertia or momentum coefficient can be introduced to limit the change in
velocity.

4.3.4 Ant Colony Optimization

The Ant Colony Optimization (ACO) algorithms are inspired by the foraging behavior of
ants. While walking from food sources to the nest and vice versa, ants deposit pheromone on the
ground. When they decide about a direction to go, with higher probability they choose paths
that are marked by stronger pheromone concentrations. This basic behavior is the basis for a
cooperative interaction which leads to the emergence of shortest paths.
Ant Colony System is the most popular ACO algorithm. Its pseudo code listing is provided
on Figure 4.5.
The probabilistic step-wise construction of solution makes use of both history (pheromone)
and problem-specific heuristic information to construct a solution piece-by-piece. Each compo-
nent can only be selected if it has not already been chosen. For components that can be selected
given current component i, the selection probability is defined as:
α ∗ ηβ
τi,j i,j
Pi,j ← Pc β
, (4.4)
α ∗ ηi,k
k=1 τi,k

where ηi,j is the maximizing contribution to the overall score of selecting the component, β is
the heuristic coefficient (commonly fixed to 1.0), τi,j is the pheromone value for the component,

29
4.3. METAHEURISTICS

α is the history coefficient, and c is the set of usable components. Greediness factor q 0 is used
to decide when to use the above probabilistic component selection and when to greedily select
the best possible component.
A local pheromone update is performed for each solution that is constructed to dissuade
following solutions to use the same components in the same order, as follows:

0
τi,j ← (1 − σ) ∗ τi,j + σ ∗ τi,j , (4.5)
where τi,j represents the pheromone for the component (i, j), σ is the local pheromone factor
0 is the initial pheromone value.
and τi,j
At the end of each iteration, the pheromone is updated and decayed using the best candidate
solution found so far as follows:

τi,j ← (1 − ρ) ∗ τi,j + ρ ∗ ∆τ i, j, (4.6)

where τi,j represents the pheromone for the component (i, j), ρ is the decay factor, and ∆τ i, j
is the maximizing solution cost for the best solution found so far if the component ij is used in
the globally best known solution, otherwise it is 0.

Figure 4.5: Pseudo code for Ant Colony System [21]

The local pheromone coefficient σ controls the amount of contribution history plays in a
components probability of selection and is commonly set to 0.1. The heuristic coefficient β
controls the amount of contribution problem-specific heuristic information plays in a components
probability of selection and is commonly chosen between 2 and 5. The decay factor ρ controls
the rate at which historic information is lost and is commonly set to 0.1. The greediness factor
q 0 is commonly set to 0.9. The total number of ants m is commonly set low, such as 10 [21].

30
4.3. METAHEURISTICS

4.3.5 Evolutionary Algorithms

Contrary to Simulated Annealing and Tabu Search, Evolutionary Algorithms operate not
only in one solution, but in a population of solutions. The population is arbitrarily initialized,
and it evolves toward better regions of the search space by means of randomized processes of
selection, mutation and recombination. One iteration of this loop is referred to as a generation.
Each individual in the population receives a measure of its fitness in the environment. Selection
focuses attention on high fitness individuals, thus exploiting the available fitness information.
Mutation introduces innovation into the population and recombination mechanism allows mixing
of parental information while passing it to their descendants.
Three main streams of evolutionary algorithms can be identified [25]:

• Genetic Algorithms (GA), developed by J. Holland [Holland, 1975] in the United States of
America, with refinements by K. De Jong [De Jong, 1975], J. Grefenstette [Grefenstette,
1986], [Grefenstette, 1987] and D. Goldberg [Goldberg, 1989]

• Evolutionary Programming (EP), originally developed by L. J. Fogel, A. J. Owens, and

M. J. Walsh in the United States of America [Fogel, Owens and Walsh, 1966] and recently
refined by D. B. Fogel [Fogel, 1991b]

• Evolution Strategies (ES), developed in Germany by I. Rechenberg [Rechenberg 1965;

Rechenberg 1973] and H.-P. Schwefel [Schwefel, 1977], [Schwefel, 1981]

Although simplistic from a biologist0 s point of view, these algorithms are sufficiently complex
to provide robust and powerful adaptive search mechanisms.

4.3.5.1 Genetic Algorithm

The Genetic Algorithm is inspired by population genetics and evolution at the population
level. Individuals of a population contribute their genotypes (genetic material) proportional
to the suitability to the environment of their phenotype (expressed genome), in the form of
offspring. The next generation is created through mating that involves recombination of two
individuals0 genomes with the introduction of mutation (random copying errors). This iterative
process may result in an improved adaptive-fit between the phenotypes of individuals in a
population and the environment.
Genetic Algorithm is implemented as following:

1. Randomly generate the first population.

2. Do until the termination criterion is satisfied:

(a) Evaluate the fitness of all of the individuals in the population.

(b) Create a new population by performing operations such as crossover, fitness-proportionate
reproduction and mutation on the individuals whose fitness has just been measured.
(c) Discard the old population and iterate using the new population.

31
4.3. METAHEURISTICS

The pseudo code listing of Genetic Algorithm is provided on Figure 4.6.

Figure 4.6: Pseudo code for Genetic Algorithm [21]

The size of the population must be large enough to provide sufficient coverage of the domain
and mixing of the useful sub-components of the solution [Goldberg, 1992]. The Genetic Algo-
rithm is classically configured with a high probability of recombination (95%-99% of the selected
population) and a low probability of mutation (1/L, where L is the number of components in
a solution) [Muhlenbein, 1992] [Back, 1993]. The fitness-proportionate selection of candidate
solutions to contribute to the next generation should be neither too greedy (to avoid the takeover
of fitter candidate solutions) nor too random.

4.3.5.2 Evolutionary Programming

Evolutionary Programming, as well as Evolution Strategies described below, is inspired by
the theory of evolution by means of natural selection. Specifically, the technique is inspired by
the species-level process of evolution (phenotype, hereditary, variation) and is not concerned
with the genetic mechanisms of evolution (genome, chromosomes, genes).
Evolutionary Programming is implemented as following:

1. Choose an initial population of trial solutions at random.

2. Replicate each solution into a new population.

3. Mutate each of these offspring solutions according to a distribution of mutation types.

4. Assess each offspring solution by computing its fitness.

32
4.3. METAHEURISTICS

The pseudo code listing of Evolutionary Programming is provided on Figure 4.7.

Figure 4.7: Pseudo code for Evolutionary Programming [21]

The sample size (BoutSize) for tournament selection during competition is commonly chosen
between 5% and 10% of the population size. There is no requirement neither that the population
size has to be held constant, nor that only a single offspring can be generated from each parent.
Typically, Evolutionary Programming does not use any crossover as a genetic operator and uses
stochastic tournament to determine solutions to be chosen for the next population.

4.3.5.3 Evolution Strategies

Schwefel generalized evolution strategies into two types: plus-strategy, denoted by (µ + λ)
and comma-strategy, denoted by (µ, λ) [26]. Both of them imitate the following basic principle:
a population, leading to the possibility of recombination with random mating, mutation and
selection. In the plus case, the parental generation is taken into account during selection, while
in the comma case only the offspring undergoes selection, and the parents die off. µ denotes the
population size, and λ denotes the number of offspring generated per generation.

33
4.3. METAHEURISTICS

The pseudo code listing of (µ, λ) Evolution Strategy is provided on Figure 4.8.

Figure 4.8: Pseudo code for (µ, λ) Evolution Strategy [21]

Evolution Strategies are commonly configured such that 1 ≤ µ ≤ λ. The ratio of µ to λ

influences the amount of selection pressure (greediness) exerted by the algorithm. The comma-
selection variation of the algorithm can be good for dynamic problem instances given its capa-
bility for continued exploration of the search space, whereas the plus-selection variation can be
good for refinement and convergence.

Each of considered algorithms emphasizes different features as being most important for a
successful evolution process.
Both ES and EP are concentrated on mutation as the main search operator, while the role of
mutation in canonical GA is usually seen to be of secondary (if any) importance. On the other
hand, recombination that plays a major role in canonical GA, is missing completely in EP, and
is urgently necessary for use in connection to self-adaptation in ES.
Finally, both canonical GA and EP emphasize a necessarily probabilistic selection mecha-
nism, while from the ES point of view selection is completely deterministic without any evidence
for the necessity of incorporating probabilistic rules. In contrast, both ES and EP definitely ex-
clude some individuals from being selected for reproduction, i.e., they use extinctive selection
mechanisms, while canonical GAs generally assign a nonzero selection probability to each indi-
vidual.
In current thesis, Evolutionary Algorithms and, specifically, Evolution Strategies were chosen
for solving Vehicle Routing Problem with Time Windows. There are several reasons for that
decision. The main reason is that Evolutionary Algorithms operate on the big search space;
they are working not only with one solution, but with a population of solutions. Furthermore it
is possible to introduce changes in formulation of the problem, mute on new parameters into the
algorithm and add new constraints. The last but not least reason is the possibility of parallel
computing with the usage of clusters.

34
Chapter 5

Decision Support Tool

5.1 Algorithm Outline

As it was mentioned before, in this thesis ground handling staff scheduling problem is pre-
sented as multiple Vehicle Routing Problems with Time Windows (VRPTW), identified for each
worker qualification.
VRPTW is a combinatorial optimization and integer programming problem, the solution of
which is the optimal set of routes for a fleet of vehicles to traverse in order to deliver services to
a given set of customers within the specified time windows.
Ground handling agents are considered as vehicles; tasks that need to be done during
aircrafts0 turnaround are considered as customers.
Evolutionary Strategies were chosen to look for a solution of VRPTW. The reasons for this
choice were given in the previous chapter.

Basic steps of ES algorithm are the following:

1. Initialize individuals (candidate solutions) to form the initial parent population

2. Repeat until termination criterion is satisfied:

(a) Select mating pool from the parent population to generate offsprings
(b) Mutate children
(c) Select the best individuals from the population of children and parents based on the
values of fitness function
(d) Use the selected individuals as parent population for the next generation.

In Evolutionary Strategies, population is the list of individuals, which are represented by

their phenotype, containing the candidate values of the parameters being optimized. Therefore
in VRPTW for ground handling, individuals are the possible set of routes, which consist of
sequences of customers for each vehicle.

Two case studies corresponding to two different implementations of algorithm will be con-
sidered in this thesis. The main difference of case study 2 compared to case study 1 is that all
the individuals generated initially or obtained as a result of applying genetic operators represent
feasible solutions, i. e. they satisfy all constraints described in section 3.6.

35
5.2. INITIAL POPULATION

5.2 Initial Population

5.2.1 Case 1
The first population is generated randomly. This means, that which customers, how many
of them and in which order they have to be visited by each vehicle are chosen at random, while
making sure that all the customers are visited exactly one time. This procedure has to be done
for each individual in first population.
It is also essential to check that worker is able to perform all tasks assigned to him. As for
example, pushback operation has to be done by a driver, though it cannot be assigned to the
driver that can only perform bus driving.

5.2.2 Case 2
The initial population is no longer generated randomly (generateInitialPopulation(...), Annex
A). The route for the first vehicle is being created in the following way: the customer is added
to the end of the current route if and only if both time windows constraints and working hours
constraints as well as the ability to perform this particular task are satisfied (CheckTAS(...),
Annex A).. This procedure is repeated till there are no more customers that could be added
to the route. Then the same idea is applied to construct routes for other vehicles, though only
unvisited yet customers are considered. Each individual in population is generated in described
way. In this case the feasibility of initial solutions is kept.

5.3 Termination Criterion

In this work, as a termination criterion was chosen the number of iterations. The another
idea would be to use the difference in resulting values of fitness function. However, in that case
we would need to introduce changes in termination criterion each time when tuning weights in
fitness function, which will result in difficulties in comparison different runs of the algorithm.

5.4 Tournament Selection

Mating pool is of size of population and is created using Ten-Tournament selection (CreateM-
atingPool(...), Annex A). This type of selection involves running several tournaments among ten
individuals chosen at random from the population. The winner of each tournament (the one
with the best fitness) is selected for mating pool (Tournament(...), Annex A).
Child is equal to its parent, randomly chosen from mating pool, modified with some proba-
bility by genetic operations (mutated).
The new generation is made out of children and parents using elitism – individuals are sorted
by fitness, after that the best half is chosen.

5.5 Genetic Operations

5.5.1 Case 1
In this work children could undergo two mutations with different probabilities (GeneticOp-
erators(...), Annex A). Those mutations were chosen in order to maintain the genetic diversity
from one generation of population to the next one while complying with the rule that all the
customers have to be visited exactly once.

36
5.6. FITNESS FUNCTION

• Mutation 1
Changing the order of visiting customers for the same vehicle, f. e.:

route1: [2 9 10] → route1: [9 2 10]

route2: [4 6] → route2: [6 4]
route3: [3 1] → route3: [3 1]
route4: [5 7 8] → route4: [8 7 5]

Individual can undergo this mutation for several vehicles.

• Mutation 2
Swapping of parts of the routes between two different vehicles, f. e.:

route1: [9 2 10] → route1: [3 1 10]

route2: [6 4] → route2: [6 4]
route3: [3 1] → route3: [9 2]
route4: [8 7 5] → route4:[8 7 5]

As mutation 1, individual can undergo mutation 2 several times. The additional benefit
of this mutation is that part of the first route could be swapped with a zero part of the
second route (or the other way around), therefore mutation 2 will work as an insertion of
part of one route into another.

After the mutations, heuristic was added to speed up the convergence. The idea of heuristic
is to find the shortest non-zero route and add it in the end of other randomly chosen route.
This ”merging” helps to minimize the number of situations when a vehicle has to serve just 1-2
customers during the whole day.

5.5.2 Case 2
Here genetic operation replicates mutation 2 from the previous case with the difference that
mutation is only applied if it will not violate the feasibility.

5.6 Fitness Function

5.6.1 Case 1
Fitness function is the objective function described in the previous Chapter (CalculateFit-
ness(...), Annex A). Total waiting time and total traveling time are computed with the use of
function (WaitTravelTime(...), Annex A). Total number of unserviced planes are computed by
(Penalty(...), Annex A).

5.6.2 Case 2
Since individuals in population satisfy a priori all necessary constraints, fitness function is
only introduced as the weighted sum of total travelling time and total waiting time.

Above were presented the most significant properties of two implementations of the Evolution
Strategies for VRPTW, considered in this thesis.

37
Chapter 6

Case Studies

All the experiments described in this chapter were conducted on the computer with the
following properties:

Processor Intel(R) Core(TM) i5-3317U CPU @ 1.70GHz

RAM 8 GB
System 64-bit operating system, processor x64

6.1 Input Data

Semi realistic data from Barcelona El Prat Airport was used to test performance of the
algorithm in this thesis. The algorithm was implemented in Matlab, which is a great environment
to quickly prototype scientific algorithms with the benefit of being able to set a breakpoint and
then run preceding the breakpoint code and plot data for interim results. However execution
time is slow compared to almost every other programming language. That is the reason why
only partial data was used to imitate the real situation. The input consists of 4 gates, 50 flights
of 2 airlines with 2 different types of aircrafts which are scheduled from 6am till midnight and
10-12 workers of each qualification. The input details are described below.

• Gates
G = {g1 ...g4 } – set of gates
g1 and g2 are of the type finger; gates g3 and g4 are of the type remote; The difference
between those two types of gates is that in cases when the aircraft is serviced on a remote
gate, there should be provided bus services for disembarking/boarding of passengers.

0 1 2 3 1
1 0 1 2 2
D(G) = 2 1 0 1 3
3 2 1 0 4
1 2 3 4 0

D(G) – cost distance matrix between the gates. The entry (D(G))ij is the time in minutes
spent by worker to move from the gate gi to gj , i, j = 1..5.
Gate g5 corresponds to the depot {0}.
It is considered that gates are located in line and on the same distance from each other;
therefore workers need 1 minute to move from gate 1 to 2 or 2 to 3, but 2 minutes to move
from gate 1 to 3 and so on.

38
6.1. INPUT DATA

• Flights
Flight timetable is presented in the following way:

Airline Aircraft Flight STA Flight STD Gate

VY A330-200 3009 00615 3011 00700 3
LH A340-300 5035 00635 5036 00735 1
VY A330-200 200 00640 202 00725 2
LH A340-300 1269 00700 1270 00800 4
VY A330-200 3165 00715 3163 00800 3
LH A340-300 6035 00740 6036 00840 2
VY A330-200 1072 00815 1071 00900 1
VY A330-200 214 00820 215 00905 3
VY A330-200 5002 00820 5003 00905 4
...

In the table above STA stands for Scheduled Time of Arrival, which is presented in a way
”DHHMM”, where D is a day, HH refers to hours (24 hour time format) and MM denotes
minutes. STD stands for Scheduled Time of Departure and is also of type ”DHHMM”.
It is considered that turnaround time of aircrafts A330-200 operated by Vueling takes 45
min, while for A340-300 operated by Lufthansa it is 60 min.

• Workers
Working hours of ground handling agents are represented as following table:

Name Qualification Tasks Start End

Juan supervisor supervising 00600 01500
Roger supervisor supervising 01200 02100
Alberto driver bus, pushback 00600 01500
Manuel driver bus 01200 02100
Christian ramp agent disembarking, boarding 02100 10600
Andreas refueler refueling 00600 01500
Olivier baggage handler unloading, loading 01200 02100
Alexey caterer catering 02100 10600
Vladimir cleaning servicer cleaning 00600 01500
...

To each worker one of the following shifts is assigned:

Morning shift 06.00 - 15.00

Afternoon shift 12.00 - 21.00
Night shift 21.00 - 06.00

Workers of the same qualification could be able to carry out different tasks. For example,
driver Alberto can perform both bus driving and pushback operations, while driver Manuel
is only able to do the bus driving.

39
6.1. INPUT DATA

• Tasks
The list of tasks, their duration and qualification of workers needed to perform these
tasks are determined for each type of aircraft separately. For example, for A330-200 this
information is performed in the following way:

Task Qualification Duration

supervising supervisor 60
bus service for unboarding of passengers driver 18
bus service for boarding of passengers driver 20
pushback driver 2
disembarking of passengers ramp agent 8
boarding of passengers ramp agent 15
refueling refueler 13
unloading of luggage baggage handler 11
loading of luggage baggage handler 18
catering caterer 15
cleaning cleaning servicer 13

Duration of the tasks is given in minutes.

• Service Level Agreement (SLA)
Necessary information from SLA is given in the form:

Airline Aircraft Task Number of Duration Trigger Time Trigger Time

workers start end
LH A340-300 cleaning 1 18 STA 12 STD -20
LH A340-300 catering 1 20 STA 12 STD -20
LH A340-300 unloading 1 22 STA 0 STA 25
LH A340-300 loading 1 26 STA 25 STD 0
LH A340-300 pushback 1 2 STD 0 STD 2
VY A330-200 supervising 1 60 STA -5 STD 10
VY A330-200 bus 2 18 STA -5 STA 13
VY A330-200 bus 2 20 STD -20 STD 0
VY A330-200 disembarking 1 8 STA 0 STA 8
VY A330-200 boarding 1 15 STD -15 STD 0
...

Values in columns ”Duration” and ”Time” are given in minutes.

In this thesis, the following precedence restrictions are represented in the schedule through
Service Level Agreement table:
– Loading of baggage could be performed only after unloading.
– Fueling, catering and servicing could be performed only in between disembarking and
boarding of passengers.
– Bus service has to be performed during the whole process of disembarking of pas-
sengers plus some time before/after, and during the whole process of boarding of
passengers plus some time before.

There are also some parameters that have to be tuned before running the code: size of
population, maximum number of iterations, probability of choosing the best individual in
tournament selection, probabilities of mutations 1 and 2, as well as weights used in fitness
function.

40
6.2. CASE STUDY 1

6.2 Case Study 1

6.2.1 Presentation of Results
To get the full schedule for ground hanling agents Matlab code should be rerun for each
qualification of the workers. There are slight differences between formulations of VRPTW for
some of workers0 qualifications that were taking into account in programming code in this thesis.
For example, the fact that some of workers are able to perform more than one task, or the fact
that aircrafts need to be serviced by workers of some qualification twice (ramp agents performing
boarding and unboarding of passengers, bus drivers, etc.)
Performing and analysing the full schedule is a tedious task, therefore below will be presented
results achieved by running the algorithm for one of the qualifications: refuelers.

The algorithm was run with the following parameters:

P Size = 500 population size

M axIter = 1000 number of iterations
α1 = 4 weight for travel time in fitness function
α2 = 2 weight for waiting time in fitness function
α3 = 100 weight for number of planes serviced out of time windows in fitness function
α4 = 10000 weight for number of planes serviced out of shift time in fitness function
pM utation1 = 0.3 probability of mutation1
pM utation2 = 0.4 probability of mutation2

for 8 workers of a qualfication ”refueler”, which could only perform one task ”refueling” and
whose working hours are the following:

Name Qualification Tasks Start End

Juan refueler refueling 00600 01500
Alberto refueler refueling 00600 01500
Andreas refueler refueling 01200 02100
Christian refueler refueling 01200 02100
Olivier refueler refueling 01200 02100
Alexey refueler refueling 01200 02100
Vladimir refueler refueling 01900 10100
Rodrigo refueler refueling 01900 10100

The output of the model is performed as a schedule and is shown on Table 6.1:

41
6.2. CASE STUDY 1

Name Qualification Tasks Start End Flight Gate

Juan refueler refueling 0:06:23 0:06:36 3009 3
Juan refueler refueling 0:06:48 0:07:01 200 2
Juan refueler refueling 0:07:12 0:07:30 1269 4
Juan refueler refueling 0:10:32 0:10:50 1036 2
Juan refueler refueling 0:11:48 0:12:01 105 2
Juan refueler refueling 0:13:18 0:13:31 420 1
Alberto refueler refueling 0:06:47 0:07:05 5035 1
Alberto refueler refueling 0:10:57 0:11:15 2098 1
Alberto refueler refueling 0:11:17 0:11:35 100 3
Alberto refueler refueling 0:11:42 0:12:00 1038 4
Alberto refueler refueling 0:13:57 0:14:15 657 2
Andreas refueler refueling 0:15:12 0:15:30 689 1
Andreas refueler refueling 0:15:33 0:15:46 5002 4
Andreas refueler refueling 0:16:12 0:16:30 3400 3
Andreas refueler refueling 0:16:42 0:17:00 345 2
Andreas refueler refueling 0:18:08 0:18:21 1277 3
Andreas refueler refueling 0:18:42 0:19:00 7000 4
Christian refueler refueling 0:13:48 0:14:01 288 3
Christian refueler refueling 0:15:08 0:15:21 4000 3
Christian refueler refueling 0:15:21 0:15:34 3165 3
Christian refueler refueling 0:16:12 0:16:30 266 4
Christian refueler refueling 0:17:47 0:18:05 540 1
Christian refueler refueling 0:20:08 0:20:21 990 1
Christian refueler refueling 0:20:33 0:20:46 134 3
Olivier refueler refueling 0:12:08 0:12:21 7084 1
Olivier refueler refueling 0:12:23 0:12:36 214 3
Olivier refueler refueling 0:12:48 0:13:01 354 2
Olivier refueler refueling 0:15:08 0:15:21 5544 4
Olivier refueler refueling 0:19:23 0:19:36 4067 3
Olivier refueler refueling 0:20:03 0:20:16 9226 4
Olivier refueler refueling 0:20:17 0:20:35 1266 3
Alexey refueler refueling 0:12:01 0:12:14 1072 1
Alexey refueler refueling 0:13:48 0:14:01 5403 4
Alexey refueler refueling 0:16:27 0:16:45 1081 1
Alexey refueler refueling 0:17:53 0:18:06 1021 2
Alexey refueler refueling 0:18:07 0:18:25 1034 1
Alexey refueler refueling 0:18:58 0:19:11 1002 1
Vladimir refueler refueling 0:19:02 0:19:15 698 2
Vladimir refueler refueling 0:19:17 0:19:35 5039 4
Vladimir refueler refueling 0:20:07 0:20:25 1356 2
Vladimir refueler refueling 0:21:38 0:21:51 3008 4
Vladimir refueler refueling 0:21:53 0:22:06 190 2
Rodrigo refueler refueling 0:21:13 0:21:26 2587 1
Rodrigo refueler refueling 0:21:27 0:21:40 2087 2
Rodrigo refueler refueling 0:21:40 0:21:53 257 2
Rodrigo refueler refueling 0:22:18 0:22:31 1012 1
Rodrigo refueler refueling 0:22:42 0:23:00 3210 2
Rodrigo refueler refueling 0:23:01 0:23:19 1044 3
Rodrigo refueler refueling 0:23:23 0:23:36 344 1
Rodrigo refueler refueling 0:23:37 0:23:55 6035 2

Table 6.1: Output of the algorithm. Serviced in time aircrafts are highlighted in green, unserviced aircrafts are
highlighted in red
42
6.2. CASE STUDY 1

As it could be seen from Table 6.1 algorithm does not find a feasible solution: 11 aircrafts
out of 50 stay unserviced. If the number of iterations and population size is increased (as f. e.
till 5000), the number of unserviced aircrafts is smaller but even then the algorithm cannot find
the feasible solution. In case of big number of iterations and population size, the computational
time will increase to many hours, which will kill the utility of algorithm in real life.
Figure 6.1 shows tendency of fitness function of best individual in population on each iter-
ation. The value of fitness function decreases but in much slower tempo when increasing the
number of iterations.

Figure 6.1: Fitness function of best individuals on each population

From Figure 6.2 and Figure 6.3 could be seen that the algorithm immediately find solutions
with zero P enalty1 - number of planes scheduled to be serviced out of shift time, but looking for
solutions with zero P enalty2 - number of planes scheduled to be serviced out of required time
windows goes much slower. That make sense since at the beginning, the priority which one to
minimize was given to P enalty1 using weight coefficients in fitness function.

Figure 6.2: Number of planes scheduled to be Figure 6.3: Number of planes scheduled to be
serviced out of workers0 shift serviced out of required time windows

43
6.2. CASE STUDY 1

6.2.2 Computational Time Analysis

The execution time of the algorithm run for workers of qualification ”refueler” with the
parameters described above is 11,4 minutes.

In order to measure on which functions the code spends the most time Profiler - graphical
user interface provided by Matlab - was used. Results, from which it is easy to see the most
time-consuming functions are shown on Figure 6.4.

Figure 6.4: Part of the Profiler report

Profiler helps to identify performance problems, such as:

• Unnecessary computation arised from oversight

• Costly functions, which could probably be replaced by others

• Recomputations, that could be avoided by storing results for future use

and is used to improve general performance of the algorithm.

Figures 6.5 and 6.6 present dependance of the run time of algorithm from number of iterations
and the size of population respectively. Number of iterations that were taken: 50, 100, 250, 500,
800, 1000 and 1500, while population size was 500. Population size considered: 50, 100, 250,
500, 800 and 1000, while number of iterations was 1000.
Easy to notice that the dependance is linear. From Figure 6.4 is clear, that functions with
longer execution time, such as GeneticOperators and Calculate start end are called once on
iteration and function CalculateFitness which appeals to both Penalty and WaitTravelTime is
called (the size of population * number of iterations) times. That explains the linear dependance.

44
6.2. CASE STUDY 1

Figure 6.5: Diagram showing the relationship between run time of the algorithm measured in seconds
and number of iterations

Figure 6.6: Diagram showing the relationship between run time of the algorithm measured in seconds
and size of population

45
6.3. CASE STUDY 2

6.3 Case Study 2

6.3.1 Presentation of Model Output
In case study 2 values of the size of population and number of iterations could be taken
much smaller comparint to the case study 1 since the algorithm is already working in a space of
feasible solutions. The probability of mutation is quite high in order to introduce more diversity
into population.

The algorithm was run with the following parameters:

P Size = 100 population size

M axIter = 200 number of iterations
α1 = 4 weight for travel time in fitness function
α2 = 2 weight for waiting time in fitness function
pM utation1 = 0.8 probability of mutation1

for the same workers as in case 1 with the following working hours:

Name Qualification Tasks Start End

Juan refueler refueling 00600 01500
Alberto refueler refueling 00600 01500
Andreas refueler refueling 01200 02100
Christian refueler refueling 01200 02100
Olivier refueler refueling 01200 02100
Alexey refueler refueling 01200 02100
Vladimir refueler refueling 01900 10100
Rodrigo refueler refueling 01900 10100

The output of the model is performed as a schedule and is shown on Table 6.2.
Figure 6.7 shows tendency of fitness function of best individual in population on each itera-
tion.

Figure 6.7: Fitness function of the best individual in population

46
6.3. CASE STUDY 2

Name Qualification Tasks Start End Flight Gate

Juan refueler refueling 0:06:48 0:07:01 200 2
Juan refueler refueling 0:07:12 0:07:30 1269 4
Juan refueler refueling 0:07:31 0:07:44 3165 3
Juan refueler refueling 0:07:52 0:08:10 6035 2
Juan refueler refueling 0:08:28 0:08:41 214 3
Juan refueler refueling 0:09:47 0:10:05 1044 3
Juan refueler refueling 0:10:12 0:10:30 5039 4
Juan refueler refueling 0:10:32 0:10:50 1036 2
Juan refueler refueling 0:11:12 0:11:30 100 3
Juan refueler refueling 0:11:48 0:12:01 105 2
Juan refueler refueling 0:12:08 0:12:21 7084 1
Alberto refueler refueling 0:06:23 0:06:36 3009 3
Alberto refueler refueling 0:06:47 0:07:05 5035 1
Alberto refueler refueling 0:08:23 0:08:36 1072 1
Alberto refueler refueling 0:09:03 0:09:16 190 2
Alberto refueler refueling 0:09:42 0:10:00 1034 1
Alberto refueler refueling 0:10:57 0:11:15 2098 1
Alberto refueler refueling 0:11:42 0:12:00 1038 4
Alberto refueler refueling 0:13:18 0:13:31 420 1
Andreas refueler refueling 0:13:48 0:14:01 5403 4
Andreas refueler refueling 0:15:12 0:15:30 689 1
Andreas refueler refueling 0:15:33 0:15:46 257 2
Andreas refueler refueling 0:16:12 0:16:30 3400 3
Andreas refueler refueling 0:17:53 0:18:06 1021 2
Andreas refueler refueling 0:18:58 0:19:11 698 2
Andreas refueler refueling 0:19:23 0:19:36 4067 3
Andreas refueler refueling 0:20:07 0:20:25 1356 2
Christian refueler refueling 0:12:48 0:13:01 354 2
Christian refueler refueling 0:13:48 0:14:01 288 3
Christian refueler refueling 0:15:08 0:15:21 5544 4
Christian refueler refueling 0:16:42 0:17:00 345 2
Christian refueler refueling 0:17:47 0:18:05 540 1
Christian refueler refueling 0:18:08 0:18:21 1277 3
Christian refueler refueling 0:18:42 0:19:00 7000 4
Christian refueler refueling 0:20:03 0:20:16 9226 4
Olivier refueler refueling 0:12:37 0:12:55 1266 3
Olivier refueler refueling 0:13:57 0:14:15 657 2
Olivier refueler refueling 0:15:08 0:15:21 4000 3
Olivier refueler refueling 0:16:12 0:16:30 266 4
Olivier refueler refueling 0:16:33 0:16:51 1081 1
Olivier refueler refueling 0:18:58 0:19:11 1002 1
Olivier refueler refueling 0:20:08 0:20:21 990 1
Olivier refueler refueling 0:20:33 0:20:46 134 3
Vladimir refueler refueling 0:21:13 0:21:26 2587 1
Vladimir refueler refueling 0:21:27 0:21:40 2087 2
Vladimir refueler refueling 0:22:18 0:22:31 1012 1
Vladimir refueler refueling 0:22:42 0:23:00 3210 2
Rodrigo refueler refueling 0:21:38 0:21:51 3008 4
Rodrigo refueler refueling 0:23:23 0:23:36 344 1

Table 6.2: Output of the algorithm

47
6.3. CASE STUDY 2

6.3.2 Computational Time Analysis

The execution time of the algorithm run for workers of qualification ”refueler” with the
parameters described above is 9 minutes.

The most time-consuming functions are shown on Figure 6.8 - screenshot from Profiler report.
As in case 1 the longest takes the function GeneticOperators, though most of the time is spent
by the functions it appeals to. The function CheckTAS verifies that the time windows and
workers0 shift constraints are satisfied. It is the most important feature of the implementation
of the algorthm in case study 2, that is the reason why it is called such a big number of times.
The function generateInitialPopulation takes significant time, since the algorithm builds the first
feasible solution.

Figure 6.8: Part of the Profiler report

Figures 6.9 and 6.10 present dependance of the run time of algorithm from number of itera-
tions and the size of population respectively. Number of iterations that were taken: 50, 75, 100,
150, 200 and 250, while population size was 100. Population size considered: 50, 75, 100, 150,
200 and 250, while number of iterations was 200.

48
6.3. CASE STUDY 2

Figure 6.9: Diagram showing the relationship between run time of algorithm measured in seconds
and number of iterations

Figure 6.10: Diagram showing the relationship between run time of algorithm measured in seconds
and size of population

49
Chapter 7

Analysis of Results

In case study 1 the algorithm did not give satisfactory results, it was not able to find even the
feasible solution - schedule such that all aircrafts are serviced according to specified by Service
Level Agreement time windows.
The main reason for such a performance is that Evolution Strategies is a random search,
therefore it is in need of a large search space and big number of iterations. The drawback of
ensuring that is clearly the increased computational time which will make the algorithm no
longer beneficial for operational and real-time planning.
In this thesis, in order to improve the performance of algorithm in case study 1, the attempt
was made to use an insertion heuristic for initial population. However the experiment showed
that it does not make efficiency any better, since the high probability of applied genetic operators
renders it meaningless.

Figure 7.1: Gannt chart for the solution in case study 1

Figure 7.1 shows the solution obtained in case study 1 in the form of Gannt chart. It shows
the sequence of tasks that should be performed by each worker distributed in time.

50
 The implementation of Evolution Strategies considered in case study 2 coped well with the
posed problem. It gave the feasible schedule for ground handling agents and it took comparable
short time to finish (9 minutes). The Gannt chart for this solution is presented on Figure 7.2.

Figure 7.2: Gannt chart for the solution in case study 2

From Figure 7.1 and Figure 7.2 could be seen that in both cases shift constraints are satisfied
for all the workers. However, in case study 2 tasks are distributed more uniformly, i. e. there
are no longer big gaps in work of ground handlers. Moreover, the best solution found by the
algorithm in case study 2 does not include schedule of the agent ”Alexey”, that means that
given tasks could be performed by less number of workers than given initially providing that all
necessary constraints are satisfied. Yet is another advantage in favor of the implementation of
the algorithm in case study 2.

51
Chapter 8

Conclusions and Future work

8.1 Conclusions
All the goals stated at the beginning of this master thesis (Section 1.2) were achieved.
The first step of the work was devoted to analysis of different aspects of the ground handling
staff scheduling problem: ground handling operations, their impact on flight delays, the eco-
nomic aspects, etc. In this work, those aspects were explained in such a form that even people
who are not familiar with the topic could get through it quickly, since it is crucial for correct
understanding of the problem.
On the next step the mathematical model was formulated in the form of Vehicle Routing
Problem with Time Windows (VRPTW). By definition the problem considered in this thesis
closely fits the VRPTW problem statement, in which ground handling employees and aircrafts
refer to vehicles and customers, respectively. The aircrafts have to be serviced at gates during
time windows specified by Service Level Agreement and according to flight schedule. The al-
gorithm was implemented in Matlab in a way that its main features – objective function and
constraints – are easy to be changed.
After thorough studies of problem solving methods, the most appropriate one – Evolution
Strategies – was chosen. There are several reasons for that decision. The main reason is that
Evolutionary Algorithms operate on the big search space; they are working not only with one
solution, but simultaneously with a population of solutions. Furthermore these algorithms are
characterized by flexibility due to the possibility to introduce changes in formulation of the
problem, mute on new parameters into the algorithm and add new constraints. The last but
not least reason is the possibility of parallel computing with the usage of clusters.
The algorithm was built in Matlab and two case studies were implemented. The main
difference between them is that the algorithm in the first case explores the whole solution space
to build each solution of the next population (using two types of mutation operators), finally
acting as a random search. In the second case it looks for each next solution only in the space of
feasible solutions (using a single type of mutation). The run time of the algorithm in case 1 took
11,4 minutes to execute 1000 iterations, while the algorithm with the same input data in case
2 found better solution in 200 iterations taking 9 minutes. Comparative analysis in Chapter 7
explains advantages of the algorithm 2 implementation over the algorithm 1 in more details.
The approach proposed in this work could be an alternative option to the generic software
packages widely used nowadays, such as the one presented by INFORM. The use of generic soft-
ware has a drawback: the solution is often not fully adapted to the realities of particular ground
handling company using it. On one side, scheduling system could include ”extra” features; on
another side, it could lack some. The algorithm considered in this master thesis is flexible and
could be adapted to specifics of the company: changes in work rules, problem characteristics
and sizes of operations. It should be noticed that run-time of the algorithm can be significantly
decreased when passing the code to C++ or Java (Matlab is often used for prototyping).

52
8.2. FUTURE WORK

8.2 Future Work

There are some concepts that could be implemented in the future in order to improve the
results of the algorithm described in this master thesis. One of the ideas is to use hybridization
of Evolution Strategies with some other algorithms presented in Chapter 4. The other idea will
be the usage of local heuristics, f.e. in creation of initial population or in genetic operations.
With an eye to reduce the runtime, the same algorithm could be implemented in another
programming languge, such as C++ or Java. Since Evolution Strategies are using populations
of independent individuals, parallel computing on different clusters could be helpful to make
computations for each individuals, which could help to save some of the computational time.
Full realistic data from different ground handling companies could be collected in order to
fully analyse the performance of the proposed algorithm and measure its utility in operational
efficiency of ground handlers.

53
Bibliography

[1] Official Journal of the European Communities L 272, 25/10/1996, p. 36 – 45

[2] Slate Magazine: ”The Southwest Secret”, by Seth Stevenson, June 12 2012

[3] Eurocontrol: ”All-Causes Delay to Air Transport in Europe”, April 2015

[4] United States Department of Transportation, Bureau of Transportation Statistics

[5] Boeing 777 airplane characteristics for airport planning

[6] Yeung S. S. M., Yu I. T. S., Hui K. Y. L.: ”World at work: Aircraft cabin cleaning”,
Occupational and Environmental Medicine, 2005, p. 58 – 60

[7] Kazda A., Caves R. E.: ”Airport Design and Operation”, 2nd edition, Elsevier, Oxford,
2007, p. 183 – 184

[8] Association of European Airlines: ”Recommendations for De-Icing/Anti-Icing of Aircraft

on the Ground”, 19th Edition, 2005

[9] Marelli, S., Mattocks, G., Merry, R.:”The role of computer simulation in reducing airplane
turn time”, AERO Magazine 1, 1998

[10] Ashford N., Stanton H. P. M., Moore C. A.: ”Airport Operation”, 2nd edition

[11] Distribution Lab Analysis, Eurocontrol

[12] Trabelsi S. F., Cosenza C. A. N., Cruz L. G. Z., Mora-Camino F., ICIEOM CIO: ”An oper-
ational approach for ground handling management at airports with imperfect information”,
2013

[13] Eurocontrol: ”Airport CDM Applications. Guide”, July 2003, p. 15 – 17

[14] SESAR factsheet N01/2011

[15] Eurocontrol: NM B2B Web Services. Web interface for system-to-system interoperability,
January 15

[16] Nasser A. El-Sherbeny: ”Vehicle routing with time windows: An overview of exact, heuristic
and metaheuristic methods”, Journal of King Saud University (Science), 2010, p. 123 – 131

[17] Jayaraman Valadi, Patrick Siarry: ”Applications of Metaheuristics in Process Engineering”,

2014

[18] Bianchi L., Dorigo M., Gambardella L. M., Gutjahr W. J.: ”A survey on metaheuristics
for stochastic combinatorial optimization”, Natural Computing: an international journal 8
(2), 2009, p. 239 – 287.

[19] Kirkpatrick S., Gelatt C. D., Vecchi M. P.: ”Optimization by simulated annealing”, Science,
May 1983, p. 671 – 680

54
BIBLIOGRAPHY

[20] Cerny V.: ”A thermodynamical approach to the travelling salesman problem: An efficient
simulation algorithm”, Journal of Optimization Theory and Applications, 1985, p. 41 – 51

[21] Jason Brownlee: ”Clever Algorithms: Nature-Inspired Programming Recipes”, 2014

[22] Glover F.: ”Heuristics for Integer Programming Using Surrogate Constraints. Decision
Sciences”, 1977, p. 156 – 166

[23] Glover F.: ”Future paths for integer programming and links to artificial intelligence”, Comp.
Oper. Res., 1986, p 533 – 549

[24] Dreo J., Petrowski A., Siarry P., Taillard E.: ”Metaheuristics for Hard Optimization: Meth-
ods and Case Studies”, 2003

[25] Baeck T., Schwefel H.-P.: ”An Overview of Evolutionary Algorithms for Parameter Opti-
mization”, Evolutionary Computation Volume 1, Number 1

[26] Schwefel H.-P.: ”Numerische Optimierung von Computer-Modellen mittels der Evolution-
sstrategie”, 26. Birkhäuser, Basel, 1977

55
Annex A

56
 Archive(MaxIter)=struct('individual',[],'fitness',[],'waiting_time',[],
'travel_time',[]);
[Fitness_sort_C, I] = sort(Fitness);
P_best(1) = P(I(1));
Archive(1) = P_best(1);

%%% iterations
k=2;
for Niter = 1:(MaxIter-1)
[MatePool] = CreateMatingPool(PSize, P, pTournament, TSize);
[C] = GeneticOperators(PSize, Nveh, pMutation1, depot_gate, dist_gate,
CurrentTasks, CurrentWorkers, MatePool);
[C] = Calculate_start_end(C, PSize, Nveh, depot_gate, dist_gate,
CurrentTasks, CurrentWorkers);
for i=1:PSize
Individual = C(i).individual;
[Waiting_Time(i), Travel_Time(i), Fitness(i)] =
CalculateFitness(Individual, Nveh, depot_gate, wWait, wTravel,dist_gate,
CurrentTasks, CurrentWorkers);
Fitness_C(i) = Fitness(i);
C(i).fitness = Fitness(i);
C(i).waiting_time = Waiting_Time(i);
C(i).travel_time = Travel_Time(i);
Fitness(PSize+i) = P(i).fitness;
end
Wait_all = [Wait_all, Waiting_Time(1:PSize)];
Travel_all = [Travel_all, Travel_Time(1:PSize)];

%%% updating P_best and Archive

[Fitness_sort_C, I] = sort(Fitness_C);
P_best(k) = C(I(1)); % P_best contains best individuals per population
if Fitness_sort_C(1) < Archive(k-1).fitness
Archive(k) = P_best(k);%Archive contains best individuals over
evolution
k = k+1;
else
Archive(k) = Archive(k-1);
k = k+1;
end

%%% Merge P with C and take PSize of best, save them in P_new
P_new(PSize) = struct('individual',[],'fitness',[],'waiting_time',[],
'travel_time',[]);

%%% Elitism mechanism: Selection of P based on Fitness of C and P

[Fitness_sort, J] = sort(Fitness);

%%%If the algorithm finds a solution with penaltyShift equal to 0, the

population is restarted by duplicating this solution PSize times (but it does
it only once through all the iterations)
if J(1) <= PSize
P_temp = C(J(1));
else
P_temp = P(J(1)-PSize);
end
for i=1:PSize

57
 a = J(i);
if a <= PSize
P_new(i) = C(a);
else
a = a-PSize;
P_new(i) = P(a);
end
end
P = P_new;
%end
end
Answer1(c) = Archive(MaxIter).fitness;
Answer2(c) = Archive(MaxIter).waiting_time;
Answer3(c) = Archive(MaxIter).travel_time;
end
end

tElapsed = toc(tStart);
disp(tElapsed);
fprintf(' Fitness WaitT TravelT \n');
format short
datasave = [Answer1 Answer2 Answer3];
disp(datasave);
end

%%% Solution as timetable

Timetable(n) = Archive(Niter+1);
N = length(CurrentWorkers);
M = length(CurrentTasks);

for q=1:N
L = length(Timetable(1,n).individual(1,q).route);
for p=1:L
T{s,1} = CurrentWorkers(1,q).name;
T{s,2} = CurrentWorkers(1,q).qlf;
T{s,4} = ConvertToTime(Timetable(1,n).individual(1,q).start_time(p));
T{s,5} = ConvertToTime(Timetable(1,n).individual(1,q).end_time(p));
for m = 1:M
if Timetable(1,n).individual(1,q).route(p) == m
T{s,3} = CurrentTasks(1,m).task;
T{s,6} = CurrentTasks(1,m).flight;
T{s,7} = CurrentTasks(1,m).gate;
s = s+1;
end
end
end
end
s=1;
TT = dataset({T 'Name','Qualification', 'Task', 'Start', 'End', 'Flight',
'Gate'})
profile viewer
p = profile('info');

58
function [type_gate, dist_gate, workers, flights, SLA] = getInput(G,W,P,A)

% Saving types of the gates 0 - finger, 1 – remote

fileID = fopen('gate.txt','r');
type_gate = fscanf(fileID,'%d', G);
fclose(fileID);
% Saving distances between gates as symmetrical matrix, last row/column are
distances to depot
fileID = fopen('gate_dist.txt','r');
dist_gate = fscanf(fileID,'%d', [G+1 G+1]);
fclose(fileID);
% Saving info about workers in a structure
[name, qlf, task, shiftstart, shiftend]=textread('shifts.txt','%s%s%s%s%s',
W);
workers=struct('name',name,'qlf',qlf,'task',task,'shiftstart',shiftstart,
'shiftend',shiftend);
for i=1:W
a = workers(i).shiftstart;
[workers(i).shiftstart] = ConvertToMinutes(a);
b = workers(i).shiftend;
[workers(i).shiftend] = ConvertToMinutes(b);
c = workers(i).task;
[cc, ccc] = strtok(c,',');
ccc = {ccc(2:end)};
[workers(i).task] = [cc,ccc];
end

% Saving info about planes in a structure

[airline, type_plane, flight_arr, STA, flight_dep, STD, gate_number] =
textread('flights.txt', '%s %s %s %s %s %s %n', P);
flights=struct('airline',airline,'type_plane',type_plane,'flight_arr',flight_a
rr,'STA',STA,'flight_dep',flight_dep,'STD',STD,'gate_number',gate_number);
for i=1:P
a = flights(i).STA;
[flights(i).STA] = ConvertToMinutes(a);
b = flights(i).STD;
[flights(i).STD] = ConvertToMinutes(b);
flights(i).gate_number = gate_number(i);
end

% Saving info about SLA in a structure

[airline, type_plane, task, number_workers, duration, initial_start_trigger,
initial_start_time, initial_end_trigger, initial_end_time] =
textread('SLA.txt', '%s %s %s %n %n %s %n %s %n');
SLA=struct('airline',airline,'type_plane',type_plane,'task',task,'number_worke
rs',number_workers,'duration',duration,'initial_start_trigger',initial_start_t
rigger,'initial_start_time',initial_start_time,'initial_end_trigger',initial_e
nd_trigger,'initial_end_time',initial_end_time);
for i=1:A
SLA(i).initial_start_time = initial_start_time(i);
SLA(i).initial_end_time = initial_end_time(i);
SLA(i).number_workers = number_workers(i);
SLA(i).duration = duration(i);
end

59
end
function [workload] = estimateWorkload(P,W,A,flights,workers,type_gate,SLA)

j=1;
workload=struct('task',{},'qlf',{},'number_workers',{},'flight',{},'duration',
{},'op_start',{},'op_end',{},'gate',{});
for k = 1:P
current_STA = flights(k).STA;
current_STD = flights(k).STD;
current_type_gate = type_gate(flights(k).gate_number);
for i = 1:A
if (strcmp(flights(k).airline,SLA(i).airline) == 1) &&
(strcmp(flights(k).type_plane,SLA(i).type_plane) == 1)
% if gate is finger, not reading bus drivers
if (current_type_gate == 0) && (strcmp(SLA(i).task,'bus') == 0)
workload(j).task = SLA(i).task;
for n=1:W
if (strcmp(workload(j).task,workers(n).task{1}) == 1) ||
(strcmp(workload(j).task,workers(n).task{2}) == 1)
workload(j).qlf = workers(n).qlf;
end
end
workload(j).number_workers = SLA(i).number_workers;
workload(j).flight = flights(k).flight_arr;
workload(j).duration = SLA(i).duration;
workload(j).gate = flights(k).gate_number;
if (strcmp('STA',SLA(i).initial_start_trigger) == 1)
workload(j).op_start = current_STA + SLA(i).initial_start_time;
elseif (strcmp('STD',SLA(i).initial_start_trigger) == 1)
workload(j).op_start = current_STD + SLA(i).initial_start_time;
end
if (strcmp('STD',SLA(i).initial_end_trigger) == 1)
workload(j).op_end = current_STD + SLA(i).initial_end_time;
elseif (strcmp('STA',SLA(i).initial_end_trigger) == 1)
workload(j).op_end = current_STA + SLA(i).initial_end_time;
end
j=j+1;
% if gate is remote
elseif (current_type_gate == 1)
workload(j).task = SLA(i).task;
for n=1:W
if (strcmp(workload(j).task,workers(n).task{1}) == 1) ||
(strcmp(workload(j).task,workers(n).task{2}) == 1)
workload(j).qlf = workers(n).qlf;
end
end
workload(j).number_workers = SLA(i).number_workers;
workload(j).flight = flights(k).flight_arr;
workload(j).duration = SLA(i).duration;
workload(j).gate = flights(k).gate_number;

% calculating initial time windows

if (strcmp('STA',SLA(i).initial_start_trigger) == 1)
workload(j).op_start = current_STA + SLA(i).initial_start_time;
elseif (strcmp('STD',SLA(i).initial_start_trigger) == 1)
workload(j).op_start = current_STD + SLA(i).initial_start_time;
end

60
 % calculating initial time windows
if (strcmp('STD',SLA(i).initial_end_trigger) == 1)
workload(j).op_end = current_STD + SLA(i).initial_end_time;
elseif (strcmp('STA',SLA(i).initial_end_trigger) == 1)
workload(j).op_end = current_STA + SLA(i).initial_end_time;
end
j=j+1;

end
end
end
end
end

function [P] = generateInitialPopulation(PSize, Nveh, depot_gate, dist_gate,

CurrentTasks, CurrentWorkers)

P = struct([]);
Possible.route = struct([]);
Possible.pos_task = struct([]);
Possible.task = struct([]);
for i = 1:PSize
P(i).individual = struct([]);
for j=1:Nveh
P(i).individual(j).route = [];
P(i).individual(j).task = [];
end
Ntasks = length(CurrentTasks);
Todo = [1:Ntasks];
length_Todo = length(Todo);
Stock = [];
for j=1:Nveh
% save possible tasks
Possible.worker = j;
n = length(CurrentWorkers(j).task);
Possible.pos_task = double(CurrentWorkers(j).task{1}(1));
if n>1
for v = 2:n
if isempty(CurrentWorkers(j).task{v})==0
Possible.pos_task = [Possible.pos_task;
double(CurrentWorkers(j).task{v}(1))];
end
end
end
P(i).individual(j).pos_task = Possible.pos_task;
for k=1:length_Todo
Possible.route = [P(i).individual(j).route; Todo(k)];
% add tasks
m = length(Possible.route);
if m>0
ind = Possible.route(1);
Possible.task = double(CurrentTasks(ind).task(1));
if m>1
for u = 2:m
ind = Possible.route(u);

61
 Possible.task = [Possible.task;
double(CurrentTasks(ind).task(1))];
end
end
end
% check constraints
[ans1, Possible] = CheckTAS(Possible, depot_gate, dist_gate,
CurrentTasks, CurrentWorkers);
if ans1 == 0
P(i).individual(j).route = Possible.route;
P(i).individual(j).start_time = Possible.start_time;
P(i).individual(j).end_time = Possible.end_time;
P(i).individual(j).task = Possible.task;
else
Stock = [Stock; Todo(k)];
end
end
Todo = Stock;
length_Todo = length(Todo);
Stock = [];
Possible.route = struct([]);
Possible.pos_task = struct([]);
end
if isempty(Todo) ~= 0
msg = 'Error: too many tasks.';
error(msg)
end
end
end

function [ans1, Current] = CheckTAS(Current, depot_gate, dist_gate,

CurrentTasks, CurrentWorkers)

ans1 = 0;
m = length(Current.route);
i = Current.worker;
[Current] = Calculate_start_end_2(Current, depot_gate, dist_gate,
CurrentTasks, CurrentWorkers);
if m>0
current_dist = dist_gate(CurrentTasks(Current.route(m)).gate,
depot_gate); % time from gate to depot
current_late_end = CurrentWorkers(i).shiftend - current_dist; % time
worker should leave gate for depot
for j = 1:m
if Current.end_time(j) > CurrentTasks(Current.route(j)).op_end
ans1 = ans1 + 1; % finishing after TW penalty
end
if Current.end_time(j) > current_late_end
ans1 = ans1 + 1; % finishing after shift ends penalty
end
end
h = ismembc(Current.task, sort(Current.pos_task));
if ~all(h) == 1
ans1 = ans1 + 1; % not being able to do this task penalty
end
end
end

62
function [P] = Calculate_start_end(P, PSize, Nveh, depot_gate, dist_gate,
CurrentTasks, CurrentWorkers)

for k = 1:PSize % going through whole population

for i = 1:Nveh % going through all vehicles
Current_route = P(k).individual(i).route;
P(k).individual(i).start_time = [];
P(k).individual(i).end_time = [];
m = length(Current_route);
if m>0
for j = 1:m
if j>1
current_end_time_prev = End_time(j-1);
current_dist = dist_gate(CurrentTasks(Current_route(j-
1)).gate,CurrentTasks(Current_route(j)).gate);
else
current_end_time_prev = CurrentWorkers(i).shiftstart;
current_dist =
dist_gate(depot_gate,CurrentTasks(Current_route(j)).gate);
end
current_duration = CurrentTasks(Current_route(j)).duration;
current_early_start = current_end_time_prev + current_dist;
current_op_start = CurrentTasks(Current_route(j)).op_start;
Start_time = max(current_early_start, current_op_start);
P(k).individual(i).start_time(j) = Start_time;
current_op_end = CurrentTasks(Current_route(j)).op_end;
End_time(j) = Start_time + current_duration;
P(k).individual(i).end_time(j) = End_time(j);
end
end
end
end
end

function [Current] = Calculate_start_end_2(Current, depot_gate, dist_gate,

CurrentTasks, CurrentWorkers)

Current.start_time = [];
Current.end_time = [];
m = length(Current.route);
i = Current.worker;
if m>0
for j = 1:m
if j>1
current_end_time_prev = End_time(j-1);
current_dist = dist_gate(CurrentTasks(Current.route(j-
1)).gate,CurrentTasks(Current.route(j)).gate);
else
current_end_time_prev = CurrentWorkers(i).shiftstart;
current_dist =
dist_gate(depot_gate,CurrentTasks(Current.route(j)).gate);
end
current_duration = CurrentTasks(Current.route(j)).duration;
current_early_start = current_end_time_prev + current_dist;
current_op_start = CurrentTasks(Current.route(j)).op_start;

63
 Start_time = max(current_early_start, current_op_start);
Current.start_time(j) = Start_time;
current_op_end = CurrentTasks(Current.route(j)).op_end;
End_time(j) = Start_time + current_duration;
Current.end_time(j) = End_time(j);
end
end

end

function [MatePool] = CreateMatingPool(PSize, P, pTournament, TSize)

MatePool(PSize) = struct('individual',[],'fitness',[],'waiting_time',[],
'travel_time',[]);
rands = zeros(TSize);
for i = 1:PSize
for j = 1:TSize
rands(j) = randi(PSize);
end;
[Winner] = Tournament(P, rands, pTournament);
MatePool(i) = Winner;
end
end
 

function [Winner] = Tournament(P, rands, pTournament)

R = length(rands);
Fitness = zeros(1,R);
for i=1:R
Fitness(i) = P(rands(i)).fitness;
end
rand3 = rand;
[~, J] = sort(Fitness);
if rand3<pTournament
Winner = P(J(1));
else
Winner = P(J(2));
end
end
 

64
function [C] = GeneticOperators(PSize, Nveh, pMutation1, depot_gate,
dist_gate, CurrentTasks, CurrentWorkers, MatePool)
%Mutation = swapping of customers between routes or inserting customer to
%route of other vehicle

C = MatePool;

for i = 1:PSize
% swapping of customers between routes
for j=1:Nveh
randM = rand;
if randM <= pMutation1
Current = C(i).individual(j);
Current.worker = j;
m = length(Current.route);

if m>0
rand1 = randi(m);
rand2 = randi(Nveh);
while rand2 == j;
rand2 = randi(Nveh);
end
a = Current.route(rand1);
aa = Current.task(rand1);
Current2 = C(i).individual(rand2);
Current2.worker = rand2;
n = length(Current2.route);

iter = randperm(n,n); % full random string

for uu=1:n
u = iter(uu);
b = Current2.route(u);
bb = Current2.task(u);
Current2.route(u) = a;
Current2.task(u) = aa;
Current.route(rand1) = b;
Current.task(rand1) = bb;
[ans1, Current] = CheckTAS(Current, depot_gate,
dist_gate, CurrentTasks, CurrentWorkers);
[ans2, Current2] = CheckTAS(Current2, depot_gate,
dist_gate, CurrentTasks, CurrentWorkers);
if (ans1 == 0) && (ans2 == 0)
C(i).individual(j).route = Current.route;
C(i).individual(j).task = Current.task;
C(i).individual(rand2).route = Current2.route;
C(i).individual(rand2).task = Current2.task;
break;
else
Current.route = C(i).individual(j).route;
Current.task = C(i).individual(j).task;
Current2.route = C(i).individual(rand2).route;
Current2.task = C(i).individual(rand2).task;
end
end

% inserting customer to another route

65
 Current.route = Current.route([1:(rand1-1),
(rand1+1):m]);
Current.task = Current.task([1:(rand1-1),
(rand1+1):m]);
m = length(Current.route);
[~, Current] = CheckTAS(Current, depot_gate,
dist_gate, CurrentTasks, CurrentWorkers);

% adding at the beginning

Current2.route = [a, Current2.route']';
Current2.task = [aa, Current2.task']';
[ans3, Current2] = CheckTAS(Current2, depot_gate,
dist_gate, CurrentTasks, CurrentWorkers);
if ans3 == 0
C(i).individual(j).route = Current.route;
C(i).individual(j).task = Current.task;
C(i).individual(rand2).route = Current2.route;
C(i).individual(rand2).task = Current2.task;
break;
else
Current2.route =
C(i).individual(rand2).route;
Current2.task = C(i).individual(rand2).task;
end

% adding in between
if n>1
iter2 = randperm(n-1,n-1); % full random string
for vv=1:(n-1)
v = iter2(vv);
Current2.route = [Current2.route(1:v)', a,
Current2.route((v+1):n)']';
Current2.task = [Current2.task(1:v)', aa,
Current2.task((v+1):n)']';
[ans4, Current2] = CheckTAS(Current2, depot_gate,
dist_gate, CurrentTasks, CurrentWorkers);
if ans4 == 0
C(i).individual(j).route = Current.route;
C(i).individual(j).task = Current.task;
C(i).individual(rand2).route = Current2.route;
C(i).individual(rand2).task = Current2.task;
break;
else
Current2.route =
C(i).individual(rand2).route;
Current2.task = C(i).individual(rand2).task;
end
end
end

% adding in the end

Current2.route = [Current2.route', a]';
Current2.task = [Current2.task', aa]';
[ans5, Current2] = CheckTAS(Current2, depot_gate,
dist_gate, CurrentTasks, CurrentWorkers);
if ans5 == 0

66
 C(i).individual(j).route = Current.route;
C(i).individual(j).task = Current.task;
C(i).individual(rand2).route = Current2.route;
C(i).individual(rand2).task = Current2.task;
break;
else
Current2.route = C(i).individual(rand2).route;
Current2.task = C(i).individual(rand2).task;
end

Current.route = C(i).individual(j).route;
Current.task = C(i).individual(j).task;
m = length(Current.route);
end
end
end
end
end

function [Waiting_Time, Travel_Time, Fitness] = CalculateFitness (Individual,

Nveh, depot_gate, wWait, wTravel, dist_gate, CurrentTasks, CurrentWorkers)

[Waiting_Time, Travel_Time] = WaitTravelTime(Individual, Nveh, depot_gate,

dist_gate, CurrentTasks, CurrentWorkers);
Fitness = wWait*Waiting_Time + wTravel*Travel_Time;
end
 

function [ans1, ans2] = WaitTravelTime(Individual, Nveh, depot_gate,

dist_gate, CurrentTasks, CurrentWorkers)
% ans1 = waiting time, ans2 = traveling time (max by vehicles, in numeric)

length_shift = zeros(1,Nveh);
sum_dur = zeros(1,Nveh);
Waiting_Time = zeros(1,Nveh);
Travel_Time = zeros(1,Nveh);
for i=1:Nveh
Current_route = Individual(i).route;
m = length(Current_route);
% traveling time
if m>0
Travel_Time(i) = Travel_Time(i) +
dist_gate(depot_gate,CurrentTasks(Current_route(1)).gate); % time from depot
at the beginning
Travel_Time(i) = Travel_Time(i) +
dist_gate(CurrentTasks(Current_route(m)).gate,depot_gate); % time to depot at
the end
if m>1
for j = 1:m-1
Travel_Time(i) = Travel_Time(i) +
dist_gate(CurrentTasks(Current_route(j)).gate,CurrentTasks(Current_route(j+1))
.gate);
% adding time between j and j+1
end
end

67
 end
% waiting time
length_shift(i) = CurrentWorkers(i).shiftend -
CurrentWorkers(i).shiftstart;
% length of shift
if m>0
sum_dur(i) = CurrentTasks(Current_route(1)).duration;
if m>1
for j=2:m
sum_dur(i) = sum_dur(i) +
CurrentTasks(Current_route(j)).duration;
end
end
% sum of durations of tasks
if sum_dur(i)+Travel_Time(i) < length_shift(i)
Waiting_Time(i) = length_shift(i) - sum_dur(i) - Travel_Time(i);
end
%else
%Waiting_Time(i) = length_shift(i);
end
end
ans1 = max (Waiting_Time);
ans2 = max (Travel_Time);
end
function [OutTA_Time, OutShift_Time, PenaltyTA, PenaltyShift] =
Penalty(Individual, Nveh, depot_gate, dist_gate, CurrentTasks, CurrentWorkers)

OutTA_Time = 0;
PenaltyTA = 0;
OutShift_Time = 0;
PenaltyShift = 0;
for i = 1:Nveh
Current_route = Individual(i).route;
m = length(Current_route);
if m>0
current_dist = dist_gate(CurrentTasks(Current_route(m)).gate,
depot_gate); % time from gate to depot
current_late_end = CurrentWorkers(i).shiftend - current_dist; % time
worker should leave gate for depot
for j = 1:m
if Individual(i).end_time(j) >
CurrentTasks(Current_route(j)).op_end
OutTA_Time = OutTA_Time + Individual(i).end_time(j) -
CurrentTasks(Current_route(j)).op_end;
PenaltyTA = PenaltyTA + 1; % finishing after TW penalty
end
if Individual(i).end_time(j) > current_late_end
OutShift_Time = OutShift_Time + Individual(i).end_time(j) -
current_late_end;
PenaltyShift = PenaltyShift + 1; % finishing after shift ends
penalty
end
end
end
end

end

68