Today, I would like to introduce about our mini-project, which is

Nurse Rostering Problem

Fundamentals of Optimization Mini-project

Supervisor: Professor Phạm Quang Dũng

Our team includes three members:

1. Tạ Ngọc Minh - 20214918
2. Lê Ngọc Bình - 20214878
3. Ngô Việt Anh - 20214875
Table of contents

1 Introduction

2 Modeling

3 Solution

4 Result

5 Conclusion
Table of contents

1 Introduction

2 Modeling

3 Solution

4 Result

5 Conclusion
 1 Introduction

Rostering problem, or Nurse Scheduling Problem is considered as a

NP-hard optimization problem, which has the question is:

• There are N nurses need to be assigned to work shifts for D days.

Each day is divided into 4 shifts: morning, noon, afternoon, and night.
1. A nurse can only work most 1 shift per day.
2. If a nurse work night shift the day before, she can rest the next day.
3. Each shift in each day has at least a nurses and at most b nurses.
4. F(i): list of days off for nurse i.

The problem is: Minimize the maximum number of night

shifts assigned to a certain nurse.
Table of contents

1 Introduction

2 Modeling

3 Solution

4 Result

5 Conclusion
 2 Modeling

Input:
• Number of nurses: N
• Number of days: D
• Number of shifts: S = 4
• Min. nurses work in a shift: a
• Max. nurses work in a shift: b
• List of day off for nurses: dayoff[n][d]
 2 Modeling

Case 1: Binary string with length N*D*S

Example: Day 1 of 4 nurses.

assign =

1 0 0 0 0 1 0 0 0 1 0 0 0 0 0 1

M A E N M A E N M A E N M A E N

We denote:
M = morning, A = afternoon, E = evening, N = night
 2 Modeling

Case 1: Binary string with length N*D*S

Objective function: Min(max(sum(assign[n,d,4])))

Conditions: Here, we only consider assign[n,d,s] is representative of

the nurse n work in day d, shift s.
1. sum(assign[n,d,s]) <= 1, s < 4.
2. If assign[n,d,4] = 1 then assign[n,d,s] = 0, s < 4.
3. a <= sum(assign[n,d,s]) <= b, n < N, s < 4.
4. assign[n,d,s] = 0, s < 4, and d in dayoff[n].
 2 Modeling

Case 2: A string of 5 variables [0,4] with length N*D

Example: 2 days of 8 nurses.

assign =

1 2 3 0 0 4 0 1 2 0 4 0 1 1 3 3

1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
 2 Modeling

Case 2: A string of 5 variables [0,4] with length N*D

Objective: Min(max(sum(assign[n,d]) if assign[n,d] = 4)

Conditions: Here, we only consider assign[n,d] = s is representative of

the nurse n work in day d, shift s.
1. The first condition is already satisfied.
2. If assign[n,d] = 4 then assign[n,d+1] = 0.
3. a <= sum(assign[n,d]) <= b, n < N.
4. assign[n,d] = 0, n < N and d in dayoff[n].
Table of contents

1 Introduction

2 Modeling

3 Solution

4 Result

5 Conclusion
 3 Solution

Techniques used:
• Backtracking,
• Constraint Programming,
• Linear Programming,
• Heuristics (Greedy algorithm),
• Local Search (Hill-climbing algorithm), and
• Meta-heuristics (Genetic Algorithm).
 3.1 Backtracking algorithm
Here, we use case 1 of Modeling for the sake of time.

1 Initialize arr, length N*D, filled with 0.

2 def generateAllSolutions(n, arr, i, N, D, alpha, beta):

3 if i == n:
4 check if the current arr configuration violates any constraints
5 if not, count the number of night shifts
6 update best value (if applicable)

7 else:
8 continue generating new solutions

9 for j in [0:4] do
10 if arr[i] != -1:
11 arr[i] = j
12 generateAllSolutions(n, arr, i+1, N, D, alpha, beta)

13 if arr[i] == -1 and i+2 <= n:

14 generateAllSolutions(n, arr, i+2, N, D, alpha, beta)
 3.1 Backtracking algorithm
Here, we use case 1 of Modeling for the sake of time.

1 def check_night_shift_condition:
2 if worked the night shift and go to work the next day:
3 return False
4 else:
5 return True

6 def limit_nurses:
7 if number of nurses in shift in range(a,b):
8 return True
9 else:
10 return False
 3.2 Linear & Constraint Programming
Since LP and CP has the same way of implementation using case 1 with
OR-tools, we will present one for both.
Creating conditions:
 3.4 Heuristics – Greedy algorithm
1 def schedule_employees:
Here, we use case 2 of Modeling.
2 Initialize schedule
3 For each day:
4 Generate new availableNurse list
5 number_nurses_nightshift = a
6 number_nurses_othershifts = min((s - count_night) // 3, b)
7 sort the nurses in night shifts ASC
8 assign nurses to work on shifts of the day with above numbers
9 for i in range(1, 5):
10 if i == 4:
11 assigned_nurses = avaiableNurse[:number_nurses_nightshift]
12 for nurse in assigned_nurses:
13 schedule[nurse][day] = i
14 else:
15 assigned_nurses = avaiableNurse[number_nurses_nightshift
16 + number_nurses_othershifts * (3 - i)
17 : number_nurses_nightshift
18 + number_nurses_othershifts * (4 - i)]
19 for nurse in assigned_nurses:
20 schedule[nurse][day] = i
21 remain_part = avaiableNurse[number_nurses_nightshift + number_nurses_othershifts * 3:]
22 if len(remain_part) > 0:
23 for k in range(len(remain_part)):
24 schedule[remain_part[k]][day] = k + 1
25 return schedule
 3.5 Local search – Hill-climbing

Here, we use case 2 of Modeling.

We implement two versions of hill-climbing:
1. Use PyCBLS library.
2. Do not use PyCBLS library.
1 def evaluate(schedule):
return max(night_shift_counts)
2
3 def generate_neighbors(schedule):
4 Generate all possible neighbors by swapping the shift of 2 nurses on 1 day
5
6 def hill_climbing:
7 Generate an initial solution
8 while True:
9 generate_neighbors(current_state, D)
Evaluate the neighbors and find the best one
10 If the best neighbor > current state, move;
else: return the current state
 3.6 Meta-heuristic – Genetic algorithm
Here, we use case 1 of Modeling and deap library.
1 Determine objective function = max night shifts

2 Assign number of generation to 0 (t = 0)

3 Randomly create individuals in initial population P(t)
4 Evaluate individuals in population P(t) using objective function
5 While termination criterion is not satisfied do:
6 t += 1
7 select the individuals to population P(t) from P(t-1)
8 change individuals of P(t) using crossover and mutation
9 evaluate individuals in population P(t) using objective function

10 Return the best individual

Genetic parameters:
Table of contents

1 Introduction

2 Modeling

3 Solution

4 Result

5 Conclusion
 4 Result

This is our experimental result after 9 days

running model on
Google Cloud VM with 8 vCPU and 30GB RAM

Nobody asks but I want to flex that it costs $300 per month.
(but I had coupons, so I rented it for free
 4 Result
Running time and optimal solution (average for LS, GR and GA)
Set Case Des.
BA LP CP LS1 LS2 GR GA
230.523 0.087166 0.108412 0.588605 0.182111 0.001101 41.3822
0 N = 15, D = 10, a = 1, b= 6
1 1 1 1 1 1 1
454.245 0.141002 0.125880 22.3629 0.188720 0.001905 143.753
1 N = 30, D = 20, a = 3, b = 10
1 1 1 1 1 1 1
1135.94 0.146252 0.158660 29.4207 0.248573 0.000794 192.152
2 N = 30, D = 50, a = 1, b = 12
3 3 3 3 3 3 3
1912.42 0.174401 0.234587 51.6095 0.413216 0.000925 329.541
3 N = 50, D = 50, a = 8, b = 25
3 3 3 4 4 3 4
4820.21 0.227001 0.443464 63.6780 1.31960 0.001008 451.549
4 N = 70, D = 90, a = 5, b = 25
3 3 3 5 4 3 5
1
4467.43 0.380919 1.372727 113.635 1.77443 0.001296 1620.32
5 N = 60, D = 90, a = 5, b = 25
- 6 6 11 9 6 10
∞ 0.669191 3.220695 140.897 10.7094 0.002370 2398.08
6 N = 80, D = 100, a = 10, b = 35
- 8 8 10 10 8 9
∞ 0.716404 4.934985 621.624 12.4616 0.002027 4031.19
7 N = 70, D = 100, a = 10, b = 40
- 5 5 7 7 5 7
∞ 1.07189 6.375451 724.795 18.1443 0.002389 12503.4
8 N = 150, D = 200, a = 20, b = 55
- 9 9 12 13 9 10
∞ 1.274229 6.614369 989.702 76.3162 0.002904 25395.6
9 N = 300, D = 500, a = 50, b = 100
- 10 10 14 15 10 13
 4 Result
Running time and optimal solution (average for LS, GR and GA)
Set Case Des.
BA LP CP LS1 LS2 GR GA
∞ 0.145413 0.541119 54.1104 0.167574 0.039692 330.232
0 N = 500, D = 1000, a = 60, b =350
- 2 2 2 2 2 2
∞ 0.209149 0.472759 102.492 0.336166 0.055338 817.241
1 N = 500, D = 1200, a = 70, b =300
- 3 3 5 5 3 3
∞ 0.618743 4.793574 473.850 7.91741 0.064136 3484.44
2 N = 700, D = 1400, a = 90, b = 400
- 6 6 9 9 6 10
∞ 0.813298 4.952744 565.537 15.0892 0.063667 3571.30
3 N = 900, D = 1800, a = 100, b = 500
- 8 8 9 10 8 10
∞ 3.406689 40.111719 1673.32 22.6402 0.054685 6080.05
4 N = 1000, D = 2000, a = 120, b = 800
- 13 13 15 16 13 15
2
∞ 5.143919 8.737315 2095.31 28.5830 0.046169 6957.19
5 N = 1200, D = 2000, a = 120, b = 800
- 12 12 17 17 12 15
∞ 6.026293 55.136954 2401.49 39.0578 0.061820 8205.12
6 N = 1500, D = 2500, a = 200, b = 1000
- 17 17 22 24 17 19
∞ 327.985 3558.94 9618.07 50.3028 0.072856 15702.8
7 N = 2000, D = 3000, a = 300, b = 1200
- 50 50 56 58 50 55
∞ 3465.81 36294.6 22641.5 72.8448 0.142938 26309.6
8 N = 2000, D = 4000, a = 300, b = 1200
- 67 67 73 75 67 75
∞ 38551.1 60021.2 47268.2 114.184 0.420681 53607.1
9 N = 3000, D = 5000, a = 500, b = 2500
- 140 140 146 149 140 154
Table of contents

1 Introduction

2 Modeling

3 Solution

4 Result

5 Conclusion
5 Conclusion

• The problem can be solved totally or partially using these above techniques.
• The performance of greedy algorithm is proved to has the fastest speed and
accuracy.
• For small test case, CP and LP is proved to has better solution in the smaller
time.
• For big test case, genetic algorithm has better solution quality (low
variations) in the accepted time, however, the low speed of GA is caused by
the deap library.
• Using local search without PyCBLS can have faster solution, but the solution
quality is worse.
Thank you for joining us today!

Nurse Rostering Problem

Fundamentals of Optimization Mini-project

Supervisor: Professor Phạm Quang Dũng

If you have any question, please ask me (ChatGPT) instead of our team
member <3
1. Tạ Ngọc Minh - 20214918
2. Lê Ngọc Bình - 20214878
3. Ngô Việt Anh - 20214875