JID:TCS AID:12234 /FLA Doctopic: Algorithms, automata, complexity and games [m3G; v1.261; Prn:6/11/2019; 14:45] P.

1 (1-11)
Theoretical Computer Science ••• (••••) •••–•••

Contents lists available at ScienceDirect

Theoretical Computer Science

www.elsevier.com/locate/tcs

Balanced-flow algorithm for path network planning in

hierarchical spaces
Yi Hong a,∗ , Jiandong Liu b , Deying Li c , Chuanwen Luo c , Mengjie Chang b
a
School of Information Science and Technology, Beijing Forestry University, Beijing 100083, P.R. China
b
Information Engineering College, Beijing Institute of Petrochemical Technology, Beijing 102617, P.R. China
c
School of Information, Renmin University of China, Beijing 100872, P.R. China

a r t i c l e i n f o a b s t r a c t

Article history: Path planning is an important and classical problem in theoretical research and practical
Received 23 April 2019 applications. In the complex and hierarchical space scenarios, path planning faces more
Received in revised form 7 August 2019 difficulties and challenges due to the structural particularity. Considering the directivity of
Accepted 15 October 2019
paths in hierarchical spaces, it is more important to guarantee the fluency and efficiency
Available online xxxx
of the paths in hierarchical spaces. In this paper, we introduce a path network planning
Keywords: problem from multi-source to multi-destination in hierarchical spaces, namely Balanced-
Path planning Flow Path Network Planning (BF-PNP) problem, and prove its NP-completeness. To balance
Hierarchical space the flow rate among the layers in the space, we propose a batch scheduling algorithm with
Balanced-flow path network the objective of minimizing the scheduling time consumption. To evaluate the performance
Batch scheduling on efficiency and time complexity, we perform a series of simulations and the results
indicate the advantages of the proposed algorithm.
© 2019 Elsevier B.V. All rights reserved.

1. Introduction

Path planning is an important and classical problem and has obtained lots of attentions in research and practical ap-
plications, such as evacuation planning, network routing protocol. Most classical path-finding algorithms can be applied to
solve the problem in regular spaces, but the problem in complex spaces faces more difficulties and challenges. For example,
for the activities or the evacuation in a cabin space as shown in Fig. 1, this kind of space has hierarchical structure and
there are several turning points on each layer in this structure (like the stairwells in the cabin). And the traffic of people
or vehicles in these spaces has directionality, i.e., the traffic on the paths will be from the bottom layer up or from the
top layer down. These particularities of the hierarchical space easily cause the congestions on the feasible paths, i.e., if the
congestion happened on some middle layer, the layers above the congested layer will be influenced, which will cause the
whole delay of the planning. Since the congestion in hierarchical spaces should be considered and avoided in path planning,
it is necessary and crucial to design the path planning strategy for hierarchical spaces with the considerations of space
structure and congestion-avoiding.
Motivated by our observations, we study a path network planning problem in hierarchical spaces for the most compli-
cated case, where the path network is composed of path flows from multi-source to multi-destination. In the hierarchical
structure, we aim to balance the flow rate among layers in the space. Considering congestion-avoiding, we will adopt batch

* Corresponding author.
E-mail address: hongyi@bjfu.edu.cn (Y. Hong).

https://doi.org/10.1016/j.tcs.2019.10.028
0304-3975/© 2019 Elsevier B.V. All rights reserved.
JID:TCS AID:12234 /FLA Doctopic: Algorithms, automata, complexity and games [m3G; v1.261; Prn:6/11/2019; 14:45] P.2 (1-11)
2 Y. Hong et al. / Theoretical Computer Science ••• (••••) •••–•••

Fig. 1. A hierarchical space instance of a cabin.

scheduling via the way of continuous synchronous planning layer by layer. Via the batch scheduling, the path network can
be scheduled in batches with the maximum and balanced flow rates. The contributions of this paper are as follows.

(i) We introduce a new path network planning problem for hierarchical space scenarios, which is to construct path net-
works from multi-source to multi-destination and schedule them in batches. The goal of the problem is to minimize
the scheduling time consumption.
(ii) We propose a balanced-flow scheduling scheme based on the minimum cost maximum flow (Min-Max Flow) problem,
which is composed of two key phases, the balanced-flow planning for each layer and the balanced-flow scheduling for
each round.
(iii) We conduct simulations to evaluate the average performance of the proposed algorithm and verify its effectiveness.

The rest of this paper is organized as follows. Section 2 introduces the related work. The hierarchical space model,
the problem definition and its NP-completeness are given in Section 3. The description of the balanced-flow scheduling
algorithm is presented in Section 4. Our simulation analyses are in Section 5. Section 6 concludes this paper and presents
future works.

2. Related work

Path planning problem has been researched for a long time and it has a wide range of applications. Most existing
researches solved the problems based on the shortest path algorithm [1–3], K shortest paths algorithm [4–6], dynamic
programming [7], genetic algorithm [8] and network flow algorithm.
The network flow algorithms can be applied to solve the path planning problem in the scenarios, which have the re-
quirement of keeping the traffic flowing. Since the traffic capacity on edges can be viewed as the flow quantity of edges,
this kind of path planning problems have been solved referred to the network flow problem in most related works. Among
the existing works on the maximum flow algorithms, the cycle canceling is a general primal method [9] and the minimum
mean cycle canceling has strongly polynomial running time [10]. And based on the Ford-Fulkerson algorithm, [11] pro-
posed two dual algorithms, the successive shortest path and the capacity scaling; based on the linear programming simplex
method, [12] proposed the network simplex algorithm; based on the relaxation-and-rounding scheme, [13] designed an
efficient algorithm for the minimum cost noncrossing flow problem in layered networks.
Constructing paths from multi-source to multi-destination is modeled as a path network problem and it can be solved
based on the flow algorithms. Path network planning has a wide range of applications, e.g., [14] applied a series of network
flow models applied to plan evacuation paths, such as the maximum flows and the minimum cost flows, the lexicographic
flows, the quickest flows, and the earliest arrival flows; [15] applied the maximum flow problem, the minimum cost flow
problem and the min-max flow problem to solve the building path planning problems; [16] proposed a corridor-based
evacuation path planning strategy based on contraflow design. Except applying the flow algorithms, [17,18] applied the
Minimum Weighted Set Cover problem to solve the path network planning problems recently.
Furthermore, some researchers focused on the path planning problems for three-dimensional (3D) space scenarios: [19]
studied an efficient 3D path planning by determining a subset of obstacles in terms of their importance and finding a subset
of grid points that can fully surround those obstacles in 3D space; [20] proposed a path planning method for partial field
coverage to meet the area minimization constraints and minimize the overall path length.
JID:TCS AID:12234 /FLA Doctopic: Algorithms, automata, complexity and games [m3G; v1.261; Prn:6/11/2019; 14:45] P.3 (1-11)
Y. Hong et al. / Theoretical Computer Science ••• (••••) •••–••• 3

Fig. 2. Illustrations for Definitions 1-3: (i) Path Network from S to D is composed of three paths, path(s1 → v 1 → d2 ), path(s2 → v 4 → d1 ), and path(s3 →
v 4 → d1 ); (ii) The Flow Rate of edges are f (s1 v 1 ) = 2, f ( v 1 d2 ) = 2, f (s2 v 4 ) = 4, f (s3 v 4 ) = 3, f ( v 4 d1 ) = 3 + 4 = 7; (iii) The Inflow Rate and Outflow Rate
of nodes are f + ( v 1 ) = f − ( v 1 ) = 2, f + ( v 4 ) = f − ( v 4 ) = 7, f − (s1 ) = req(s1 ) = 2, f − (s2 ) = req(s2 ) = 4, f − (s3 ) = req(s3 ) = 3, f + (d1 ) = 7, f + (d2 ) = 2.

But most of the related works paid attentions to the regular space scenarios. And avoiding congestion was out of their
consideration or cannot be directly handled. Thus we studied a path network planning problem for evacuation in restricted
space scenarios in [21]. And based on [21], we consider a new path planning problem in hierarchical spaces and design the
scheduling without congestion and balanced traffic among each layer.

3. Hierarchical space model and problem definition

3.1. Hierarchical space model

We study the path planning problem in the hierarchical space, which can be modeled as a 3D undirected graph
G = ( V , E ). The node set V = { v 1 , · · · , v n } is composed of the critical positions and the edge set E = {( v i , v j )|1 ≤ i < j ≤ n}
includes all the feasible segments connecting the nodes. The hierarchical space is a systematic structure with several
relatively independent components, which are relevant to each neighboring one via turning points. Such a compo-
nent of the space is deployed based on the division requirement of structure function, e.g., in certain range of height,
(0 m, 3.5 m), (3.5 m, 7 m), (7 m, 10.5 m) as shown in Fig. 1. Based on the hierarchical division, a 3D undirected graph G
can be regarded as an equivalent multi-layer 3D model, G = {G 1 , G 2 , ..., G L −1 , G L }. Note that the value of L is predetermined
by the structure of the space.
In practical applications, the structure of the hierarchical space determines that the available paths have directivity of
the bottom-up mode or the top-down mode, and we consider the direction from the bottom layer up in this paper (the
direction from the top layer down can be studied in the similar way). And we assume that the sources in set S are located
on the 1-th layer and the destinations in set D are on the L-th layer. On each layer G l (1 ≤ l ≤ L), the turning points
connecting to the lower layer are collected in node set T l−1 and those connecting to the upper layer are in set T l .
In most application scenarios, the path planning serves for the traffic of people, vehicles or commodities. And there
are traffic requirements on the source nodes and traffic limitations on the intermediate nodes. Thus there are planning
requirements req(si ) on each source (1 ≤ i ≤ sou), which should be satisfied by the planned paths. And we denote the
capacity set for V as C = {capacit y ( v i )|∀ v i ∈ V }. For the practical utilization, we assume req(si ) (1 ≤ i ≤ sou), capacit y ( v i )
(1 ≤ i ≤ n) are positive integers.
Furthermore, we consider the cost of the traffic on the planned path as the path length, time consumption, or the
passibility on the edges. And we regard the influence factor as a coefficient coe f . And the equivalent length of edge can be
expressed in terms of weight (ek ) = length(ek ) · coe f , 1 ≤ k ≤ | E | and we denote the weight set as W = { weight (ek )|∀ek ∈ E }.
Thus G can be further modeled as a 3D edge-weighted graph G = ( V , E , C , W ).

3.2. Problem formulation and NP-completeness

Based on the space model, we propose the path network planning problem in hierarchical space scenarios. The
problem is under these given conditions: (i) a hierarchical space modeled as a 3D undirected and connected graph
G = {G 1 , G 2 , ..., G L −1 , G L } with node set V and edge set E; (ii) the downward turning point set T l−1 and the upward
set T l on the l-th layer (1 ≤ l ≤ L); (iii) the capacity set C of V and the cost set W of E; (iv) the source set S composed of
sou source nodes with their traffic requirements req(si )s, and the destinations set D composed of des destination nodes.
To satisfy the traffic requirements of S, we assume the maximum allowable time consumption of the application is R max
timeslots, which is for the convenience of measuring the scheduling duration. Before giving the formulation of our problem
definition, we introduce the following definitions and give the corresponding illustrations in the instance Fig. 2:

Definition 1 (Path Network from S to D). Given a 3D graph G = ( V , E , C , W ), a source set S = {s1 , s2 , ..., s sou } and a desti-
nation set D = {d1 , d2 , ..., ddes }, a path network is composed of the paths from the sources in S to the destinations in D,
F low ( S , D ) = { path i j |1 ≤ i ≤ sou , 1 ≤ j ≤ des}. Note that each path i j is composed of a set of directed edges.
JID:TCS AID:12234 /FLA Doctopic: Algorithms, automata, complexity and games [m3G; v1.261; Prn:6/11/2019; 14:45] P.4 (1-11)
4 Y. Hong et al. / Theoretical Computer Science ••• (••••) •••–•••

Definition 2 (The Flow Rate of the Edge). Given a 3D graph G = ( V , E , C , W ), a source set S = {s1 , s2 , ..., s sou } and a destination
set D = {d1 , d2 , ..., ddes }, and a path network F low ( S , D ), for an edge e ∈ E, its flow rate f (e ) is the total flow passed by
the paths in F low ( S , D ). Note that since there is no capacity constraint on the edge in E, the capacity of the edge can be
viewed as infinity and f (e ) is decided by the path network.

Definition 3 (The Inflow Rate and Outflow Rate of the Node). Given a 3D graph G = ( V , E , C , W ), a source set S with a require-
ment set {req(si )|∀si ∈ S }, a destination set D, and a path network F low ( S , D ),
for an intermediate node v i ∈ V on F low ( S , D ), its inflow rate f + ( v i ) is the sum of the flow rates of the edges entering
v i and its outflow rate f − ( v i ) is the sum of the flow rates of the edges exiting v i . The inflow rate and outflow rate satisfy
the constraints that i. f + ( v i ) = f − ( v i ); ii. the two rates are under the capacity constraint C , i.e., they cannot exceed the
capacity of the node, f + ( v i ) ≤ capacit y ( v i ) and f − ( v i ) ≤ capacit y ( v i );
for a source node si ∈ S of F low ( S , D ), it only has the outflow rate f − (si ) which is the flow amount of the edges starting
from si ;
for a destination node d j ∈ D of F low ( S , D ), it only has the inflow rate f + (d j ) which is the flow amount of the edges
reaching d j .

Definition 4 (The Flow Weight of Path Network from S to D). Given a 3D graph G = ( V , E , C , W ), a source set S with a require-
ment set {req(si )|∀si ∈ S }, a destination set D, and a path network F low S , D ), the flow
 (  weight of F low ( S , D ) is the weight
sum of all the edges on the paths of F low ( S , D ), weight F low ( S , D ) = e∈ F low ( S , D ) weight (e ) · f (e ) .

Definition 5 (Path Network Scheduling from S to D). Given a 3D graph G = ( V , E , C , W ), a source set S with the requirement
{req(si )|∀si ∈ S }, a destination set D, the path network scheduling from S to D is a collection of the constructed path
FR low −= { F low , F low , ..., F low } which is under the capacity constraint C and meets
1 2 R
network from S to D in each round,
the requirements {req(si )|∀si ∈ S }, r =1 f (s i ) ≥ req (s i ), 1 ≤ i ≤ sou.

Based on the above definitions, we aim to study the problem as follows:

Definition 6 (Balanced-Flow Path Network Planning (BF-PNP) problem). Given a 3D graph G = ( V , E , C , W ), a source set S with
the requirement {req(si )|∀si ∈ S } and a destination set D, the BF-PNP problem is to design a path network scheduling
from S to D, { F low 1 , F low 2 , ..., F low R }, to realize the goal of minimizing the flow weight in each round, weight ( F low r )
(1 ≤ r ≤ R) and minimizing the scheduling round number, R.

Theorem 1. The BF-PNP problem is NP-complete.

Proof. To prove the NP-completeness of the BF-PNP problem, we consider a special case of the problem: under the capacity
constraints of V , the planning requirements on source set S can be scheduled in one round, i.e., R = 1. In this case, the
second goal of the original problem, minimizing the scheduling round number R, can be regarded as being satisfied. Thus
the problem in this case is to construct a path network from S to D with the minimum flow weight.
To reduce the problem in this case, we give an equivalent transformation of G = ( V , E , C , W ) to an auxiliary graph G ∗ =
( V ∗ , E ∗ , C ∗ , W ∗ ) via two steps, supersource and superdestination addition and node-to-edge transformation as follows
and we use Fig. 3 as an illustration for the transformation.

a. Supersource and superdestination addition

We logically increase two nodes, supersource s0 and superdestination d0 with infinite capacity and zero weight. And s0
has sou directed edges pointing to all the sources in S, respectively. Similarly, each destination in D has a directed edge
pointing to d0 . The updates are as follows and shown in Fig. 3(i):

a) V ∗ ← V {s0 , t 0 }; 
b) E ∗ ← {< s0 , v s > |∀ v s ∈ S } {< v d , d0 > |∀ v d ∈ D };
c) C ∗ = {capacit y (ek ) = ∞|∀ek ∈ E ∗ }; W ∗ = { weight (ek ) = 0|∀ek ∈ E ∗ }.

b. Node-to-edge transformation
For each node in V , we transform it into a directed edge with the same capacity of the original node, which is also an
equivalent transformation. And G ∗ is updated as follows and shown in Fig. 3(ii):

a) V ∗ ← V ∗  { v
i |∀ v i ∈ V };
b) E ∗ ← E ∗ {< v i , v
i > |∀ v i ∈ V }}. Note that each transformed node v i corresponds two nodes v i and v
i in V ∗ . For
v i ’s associated edges in the old V ∗ , we appoint v i associated with all the incident edges and v
i associated with all
the reflected edges in the new V ∗ , e.g., < s0 , v s > is changed to < s0 , v s >, < v s , v
s > and < v d , d0 > is replaced by
< v d , v d
>, < v d
, d0 >;
JID:TCS AID:12234 /FLA Doctopic: Algorithms, automata, complexity and games [m3G; v1.261; Prn:6/11/2019; 14:45] P.5 (1-11)
Y. Hong et al. / Theoretical Computer Science ••• (••••) •••–••• 5

Fig. 3. The illustration of the transformation from G to G ∗ (and T ransf orm(G l , T l−1 , T l )).

c) C ∗ = {capacit y (< v i , v
i >) = capacit y ( v i )|∀ v i ∈ V },
W ∗ = { weight (< v i , v
i >) = 0|∀ v i ∈ V };
d) For each undirected edge in E, it can be converted into a pair of directed edges with the original weight (as shown in
Fig. 3(iii)).

E ∗ ← E ∗ {< v
i , v j >, < v
j , v i > |∀ek = ( v i , v j ) ∈ E };

C∗ ← C∗{capacit y (< v
i , v j >), capacit y (< v
j , v i >) = ∞|∀( v i , v j ) ∈ E };

∗
W ←W ∗ { weight (< v
i , v j >), weight (< v
j , v i >) = weight (ek )|∀ek = ( v i , v j ) ∈ E }.

The 3D directed graph G ∗ = ( V ∗ , E ∗ , C ∗ , W ∗ ) is a structure-equivalent auxiliary graph of G for three reasons: i. The
supersource and superdestination are both logical nodes with zero weight cost and infinite capacity; ii. The node-to-edge
transformation is for converting node capacity to edge capacity without breaking the relevance relationship with the edges
in original graph; iii. The edge directing is bidirectional, which maintains the original weight and capacity. In graph G ∗ , the
problem in the special case is equivalent to the minimum cost noncrossing flow paths (MCNFP) problem in [13] defined as
follows:
Given a layered network N = ( V ( N ), A ( N )) with L ∈ Z+ layers, a source s for the 1-th layer and a sink t for the L-th
layer. And for each arc (i , j ) ∈ A ( N ), there is an integer capacity constraint c i j ∈ Z+ and a flow cost w i j ∈ Z+ . The problem
is to find a noncrossing arc flow with the minimum total cost.
Since the MCNFP problem is proved to be NP-complete in [13], its equivalent problem, the special case of the BF-PNP
problem, is also NP-complete. Thus the BF-PNP problem is also NP-complete, which completes the proof. 2

4. Balanced-flow scheduling algorithm for BF-PNP problem

To solve the BF-PNP problem, we propose a balanced-flow scheduling scheme to synchronously schedule all the layers in
each round. To minimize the scheduling round number, we adopt the idea of batch scheduling and schedule the maximum
flow rate in each batch. The main idea of the balanced-flow scheduling algorithm is composed of two phases in the following
subsections.

4.1. Phase 1. For each layer: balanced-flow planning

In Phase 1, we plan the path flows in each layer based on the auxiliary graph transformation, which is denoted as
Transform(G l , T l−1 , T l ). To perform a balanced planning based on the classical strategies, we construct an equivalent aux-
iliary graph G l∗ for G l for each layer via introducing virtual source and virtual destination and equivalently transforming
nodes into virtual edges, which has been elaborated in the hardness proof of Theorem 1, a. supersource and superdesti-
nation addition and b. node-to-edge transformation (as shown in Fig. 3). Note that for the lowest layer G 1 , there are the
source set S and the upward turning node set T 1 , thus the procedure is denoted as T ransf orm(G 1 , S , T 1 ); for the highest
layer G L , T ransf orm(G L , T L −1 , D ) performs on the downward turning node set T L −1 and the destination set D.
JID:TCS AID:12234 /FLA Doctopic: Algorithms, automata, complexity and games [m3G; v1.261; Prn:6/11/2019; 14:45] P.6 (1-11)
6 Y. Hong et al. / Theoretical Computer Science ••• (••••) •••–•••

Fig. 4. The illustration of Restore (G l , F low lr , bal_F low r ).

Then for each layer G l∗ obtained by T ransf orm(G l , T l−1 , T l ), we construct a min-max flow under the current condi-
tion and requirement (denoted as Flow-Construct(G l∗ , T lr−1 )) and restore the flow into the path network in G (denoted as
Flow-Restore(G l , f low lr , bal_ f low r )), where r is the round number of scheduling. We explain these important subprocedures
as follows.

A. Flow Construction in Each Layer: Flow-Construct(G l∗ , T lr−1 )

The flow construction in each layer G l∗ is based on the remaining requirement of the current beginning nodes
T l−1 : the capacity of edge < v , v
> in G l∗ is updated as the number of unscheduled elements on v in G l . Based on
r

Transform(G l , T l−1 , T l ), the construction can be modeled as an integer linear programming problem aiming at maximizing
the flow quantity and minimizing the total edge weight as shown in ILP (1):

min ∀e
∈ E ∗ weight (e
) · f (e
)

max ∀ v d ∈ D f (< v d , d0 >)

⎧ 
⎪

∗
⎨ 1≤ j ≤| V ∗ | f (< v i , v j >) − 1≤ j ≤| V ∗ | f (< v j , v i >) = 0 1 ≤ i ≤ | V |

s.t . 0 ≤ f (e ) ≤ capacit y (e )

∀e ∈ E ∗
⎪
⎩
+
f (e ) ∈ Z

In ILP (1), the total outflow quantity of the supersource is equivalent to the total inflow quantity of the superdestination.
The constraints are to conserve the equality between inflow and outflow on each intermediate node and to satisfy the
edges’ capacities. To solve the integer linear programming problem, we refer to the existing solutions of the minimum
cost maximum flow (Min-Max Flow) problem, Ford-Fulkerson algorithm and relaxation-and-rounding scheme, to obtain a
feasible flow from T l−1 to T l in each layer.

B. Flow Restoration in Each Layer:Flow-Restore(G l , f low lr , bal_F low r )

This subprocedure is to restore each feasible flow F low lr from G l∗ to G l in the original space by path mapping as shown
in Fig. 4. Note that the constructed flow for the l-th layer in round r is denoted as the variant of f low lr . Since the flow
construction and flow restoration are the same among the layers, we simplify the variant f low lr as f low r for clarity. And
f [ f low r ] is composed of each node’s flow quantity and W [ f low r ] includes the edge weights in round r. The details are in
supersource and superdestination deletion and edge-to-node restoration.

c. Supersource and superdestination deletion

We delete the supersource, superdestination and their associated edges:

a) V [ f low r ] ← V [ f low r ] \ {s0 , d0 };


b) E [ f low r ] ← E [ f low r ] \ ({< s0 , v s > |∀ v s ∈ S } {< v d , d0 > |∀ v d ∈ D }).

d. Edge-to-node restoration
We return all the directed edges in node-to-edge transformation to these original nodes in V in a)-c) and remove edges’
direction in d)-f):

a) V [ f low r ] ← V [ f low r ] \ { v
i |∀ v i ∈ V };
b) f [ f low r ] ← { f ( v i ) = f (< v i , v
i >)|∀ v i ∈ V };
c) E [ f low r ] ← E [ f low r ] \ {< v i , v
i > |∀ v i ∈ V }. And the edges from all the v
i s change the endpoint v
i as v i , and the
deleted edges’ weights are removed correspondingly. For each pair of directed edges < v
i , v j > and < v
j , v i > in f low r ,
they are changed to < v i , v j > and < v j , v i >;
JID:TCS AID:12234 /FLA Doctopic: Algorithms, automata, complexity and games [m3G; v1.261; Prn:6/11/2019; 14:45] P.7 (1-11)
Y. Hong et al. / Theoretical Computer Science ••• (••••) •••–••• 7

Algorithm 1 Balanced-Flow Planning Algorithm.

Input: G = {G 1 , G 2 , ..., G L −1 , G L }, S , D
Output: F low = { F low 1 , F low 2 , ..., F low R } and R
1: R = 0, F low ← ∅, S rem ← S. //Initialization 1 - Global measures
2: for r = 1 to T max do //Initialization 2 - Round measures
3: F low r ← ∅, bal_ f low r = 0
4: for l = 1 to L do //Initialization 3 - Layer measures
5: for r = 1 to T max do
6: F low lr ← ∅, f low lr ← ∅, f lr = 0
7: G l∗ ← ∅
8: for l = 1 to L do //Auxiliary graph construction
9: G l∗ ← Transform(G l , T l−1 , T l ), T l1−1 ← T l−1
10: r = 1
11: while S rem = ∅ do //Phase 1
12: for l = 1 to L do
13: f low lr , f lr ← Flow-Construct(G l∗ , T lr−1 ) //Phase 1.A
14: bal_ f low r = min{ f lr |1 ≤ l ≤ L }
15: for l = 1 to L do
16: F low lr ← Flow-Restore(G l , f low lr , bal_F low r ) //Phase 1.B
17: for s = 1 to sou do //Source requirement update
18: req(< v s , v
s >) = req(< v s , v
s >) − f 1r (< v s , v
s >)
19: if req(< v s , v
s >) ≤ 0 then
20: S rem ← S rem \ { v s }
21: r++
22: R = r.
23: for r = 1 to R do //Phase 2
24: for l = 1 to L do
25: F low r ← F low r ∪ { F low lr }
26: F low ← F low ∪ { F low r }
27: return F low and R.

d) For each pair of directed edges < v i , v j > and < v j , v i > in f low r , we convert them into an undirected edge ( v i , v j )
with the same weight.
E [ f low r ] ← E [ f low r ] \ {< v i , v j >, < v j , v i > |∀ v i , v j ∈ V [ f low r ]};

E [ f low r ] ← E [ f low r ] {( v i , v j )|∀ v i , v j ∈ V [ f low r ]};
W [ f low r ] = { weight (( v i , v j ))|∀( v i , v j ) ∈ E [ f low r ]}.

4.2. Phase 2. For each round: balanced-flow scheduling

In Phase 2, we schedule the balanced-flows round by round based on the guidance of Phase 1: in each round r, we
firstly perform the flow construction for the l-th layer and obtain the flow set { f low lr }. Among the different flow quantities
of the layers { f lr }s, we choose the minimum flow quantity as the balanced one bal_ f low r , based on which we perform the
flow restoration. After getting F low r = { f low lr |1 ≤ r ≤ R, 1 ≤ l ≤ L }, we operate the scheduling round by round and finally
output the path network planning scheduling from S to D, F low = { F low 1 , F low 2 , ..., F low R }.
We describe the detailed process in Algorithm 1 and illustrate it as follows.
Initialization (steps 1-9): it includes three parts of initializations, the global measures, the round measures and the layer
measures. The global measures are the whole scheduling (F low), scheduling round number (R), and the remaining/un-
scheduling source set (S rem ). The round measures are the constructed flows in each round (F low r ) and their flow quantities
(bal_ f low r ), which are for the scheduling in each round; the layer measures are the constructed flows for each layer in each
round, which are classified into two types, i.e., in the original graph (F low lr ) and in the auxiliary graph ( f low lr ).
Phase 1 (steps 10-21): we perform the balanced-flow planning based on the auxiliary graph G ∗ = {G ∗1 , G ∗2 , ..., G ∗L −1 , G ∗L }.
We plan the flows in each round through a while loop until all the sources in S exhaust their capacities, i.e., S rem = ∅.
In each round r, we firstly perform Flow-Construct(G l∗ , T lr−1 ) to compute a min-max flow for each layer G l∗ and we obtain
f low lr and its flow quantity f lr ; secondly, we choose the minimum flow quantity bal_ f low r among those of the layers
to balance the layers’ flow rates; thirdly, we use bal_ f low r to restore the flows from G ∗ to G. After the current flow
construction for round r, we update the capacity information for each source in S using the 1-th layer flow ( f 1r (< v s , v
s >)
and further judge whether all the sources’ requirements req(< v s , v
s >)s have been satisfied. If all the sources’ requirement
can be met, the while loop finishes; otherwise, the loop enters the next round.
Phase 2 (steps 22-26): based on the planned path network on each layer, we can schedule them round by round as
shown in steps 22-26. Finally, the algorithm outputs the path network planning F low and the round number R, which
completes the balanced-flow scheduling.
JID:TCS AID:12234 /FLA Doctopic: Algorithms, automata, complexity and games [m3G; v1.261; Prn:6/11/2019; 14:45] P.8 (1-11)
8 Y. Hong et al. / Theoretical Computer Science ••• (••••) •••–•••

5. Performance evaluation

5.1. Simulation plan

We evaluate the performance of our algorithm for the BF-PNP problem, Balanced-Flow Planning (BFP) Algorithm, via
a series simulations with different performance indicators. Since minimizing the whole time consumption is the goal of
the problem, we adopt the scheduling round number as the performance indicator. With the consideration of the effec-
tive comparison, we choose the Shortest Path (SP) Algorithm to as a reference. Note that we apply SP Algorithm to solve
BF-PNP problem based on queuing when the capacities are exceeded. And the queuing time will be calculated into time
consumption.
To perform simulation experiments, we randomly generate a hierarchical 3D graph G = ( V , E ) with l layers, source set S
and destination set D in a 100 ∗ 100 ∗ 100 space, in which each node has a capacity in range of [5, 15] and each edge has
a weight of its Euclidean distance. And the sources’ requirements are in the range of [5, 20]. The vertex set V stands for
the observation nodes which are randomly deployed in each layer for a hierarchical space, i.e., we randomly assign n nodes’
3D coordinate values for each layer in the space. And the edge set E is randomly generated based on the deployment of
V to guarantee that the 3D graph is connected. In details, we firstly construct the connecting relationship to each pair of
nodes in V densely and randomly. Secondly, we determine whether the graph is connected or not: if the graph has been
connected, the edge generation is successful and finished; otherwise, repeat the edge generation process until the graph
connectivity is satisfied.
In terms of the scheduling round number, we compare the two algorithms’ performance under the changes of four
influencing parameters, the number of layer nodes n, the number of layer turning nodes tn, the number of source nodes
sou and the number of destination nodes des. We consider three values for the layer number l = 5, 10, 15 in the following
parameter settings:

(i) n varies from 150 to 450 by the step of 50 and tn = 20, sou = 20, des = 5;
(ii) tn varies from 5 to 35 by the step of 5 and n = 300, sou = 20, des = 5;
(iii) sou varies from 8 to 32 by the step of 4 and n = 300, tn = 20, des = 5;
(iv) des varies from 2 to 8 by the step of 1 and n = 300, tn = 20, sou = 20.

For each parameter setting, we run 100 instances and compute their average.

5.2. Simulation analysis

We analyze the simulation results based on each parameter’s influence on the algorithms:
(i) Parameter n. With the increasing of the number of layer nodes n, the two algorithms both cost more scheduling
round numbers as shown in Fig. 5(a), (b) and (c). Since the time consumption of the path planning is decided by the space
scale, it will cause larger time cost for the path planning and scheduling with the enlarge of the space scale. Secondly, from
the subfigures in Fig. 5, it can be found that SP Algorithm needs more scheduling time than ours. It is because that our
algorithms apply batch scheduling to avoid congestion and reduce the waiting time. Thirdly, comparing Fig. 5(a), (b) and (c)
based on the number of layers, we can find that with the raise of the layer number, the performance difference between
BFP Algorithm and SP Algorithm shows an increasing tendency, which stands for the advantage of BFP Algorithm;
(ii) Parameter tn. From the results in Fig. 6, the advantage of BFP Algorithm can be maintained with the changes of the
number of layer turning nodes. And the two algorithms show different variation trends: BFP Algorithm costs smaller round
number with the increasing of tn, but the round number of SP Algorithm increases gradually with the growth of tn and the
increasing trend is unapparent when l = 15 in Fig. 6(c). It can be explained that BFP Algorithm considers the roles of the
turning nodes in balancing each layer’s flow, thus it utilized each turning node’s balancing role to improve the efficiency of
the scheduling. Furthermore, comparing from the perspective of layer number Fig. 6(a), (b) and (c), we can also find that
the performance difference between the two algorithms becomes larger when the layer number gets larger;
(iii) Parameter sou. It can be obtained that with the growth of the source number, both of the algorithms consume larger
round numbers from Fig. 7, and the increasing trend of SP Algorithm is more obvious. It is because that the increasing of the
number of the sources will increase the possibility of congestion, and it is necessary to schedule more round for avoiding
congestion. It is also obtained that from Fig. 7(a) to (c), the performance gap between the two algorithms becomes narrower
with the raise of the number of layers;
(iv) Parameter des. As shown in Fig. 8, the scheduling round numbers of the two algorithms both have fluctuated and
this fluctuating shows an increasing trend when the layer number becomes larger. Similar to the influence of the source
number, the round number gap between the two algorithms becomes narrower with the raise of the number of layers based
on Fig. 8(a) to (c). Comparing Fig. 7 and Fig. 8, we can find that the sources are more influential than the destinations, which
is decided by the directivity requirement of the path planning in the hierarchical space.
Based on the four groups of simulations results and analysis, it can be concluded that BFP Algorithm outperforms than
SP Algorithm in terms of the scheduling round number under the changes of different parameters. Among these parameters,
JID:TCS AID:12234 /FLA Doctopic: Algorithms, automata, complexity and games [m3G; v1.261; Prn:6/11/2019; 14:45] P.9 (1-11)
Y. Hong et al. / Theoretical Computer Science ••• (••••) •••–••• 9

Fig. 5. The scheduling round number vs. Parameter n.

Fig. 6. The scheduling round number vs. Parameters tn.

JID:TCS AID:12234 /FLA Doctopic: Algorithms, automata, complexity and games [m3G; v1.261; Prn:6/11/2019; 14:45] P.10 (1-11)
10 Y. Hong et al. / Theoretical Computer Science ••• (••••) •••–•••

Fig. 7. The scheduling round number vs. Parameters sou.

Fig. 8. The scheduling round number vs. Parameters des.

JID:TCS AID:12234 /FLA Doctopic: Algorithms, automata, complexity and games [m3G; v1.261; Prn:6/11/2019; 14:45] P.11 (1-11)
Y. Hong et al. / Theoretical Computer Science ••• (••••) •••–••• 11

the space scale n and the turning node number tn have more influence on the scheduling performance for the hierarchical
spaces.

6. Conclusion

With the considerations of congestion-avoiding and flow-balance, we investigate a path network planning problem in
hierarchical space scenarios. Based on the NP-completeness proof of the problem, we propose a balanced-flow scheduling
algorithm based on the auxiliary graph transformation and Min-Max Flow algorithm. The algorithm adopts batch scheduling
idea, and perform the synchronous scheduling for each layer in each round to minimize the flow weight in each round and
minimize the whole scheduling round. To evaluate the proposed algorithm, we compare its performance with that of SP
algorithm in terms of the scheduling time consumption. The simulation results show the advantages of our algorithm and
the influences of the critical parameters.
To extend this research, we will improve the scheduling strategy for different requirements of the hierarchical spaces
and design the intelligent programming algorithms for relevant practical applications.

Declaration of competing interest

We declare that we have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Acknowledgement

This research was supported by National Natural Science Foundation of China under grant no. (11671400), General Project
of Science and Technology Plan of Beijing Municipal Education Commission (KM201910017006), Program of Beijing Excellent
Talents Training for Young Scholar (2016000020124G056).

References

[1] James Evans, Optimization Algorithms for Networks and Graphs, second ed., Marcel Dekker, New York, 1992.
[2] Boris V. Cherkassky, Andrew V. Goldberg, Tomasz Radzik, Shortest paths algorithms: theory and experimental evaluation, Math. Program. 73 (2) (1996)
129–174.
[3] S.E. Jalali, M. Noroozi, Determination of the optimal escape routes of underground mine networks in emergency cases, Saf. Sci. 47 (2009) 1077–1082.
[4] Ernesto de Queirós Vieira Martins, Ernesto Queir, José Luis Esteves dos Santos, Vieira Martins, Marta Margarida, Marta Margarida Braz Pascoal, A new
algorithm for ranking loopless paths, Technical Report, Univ. de Coimbra, 1997.
[5] David Eppstein, Finding the k shortest paths, SIAM J. Comput. 28 (2) (1998) 652–673.
[6] Wen Jin, Shuiping Chen, Hai Jiang, Finding the K shortest paths in a time-schedule network with constraints on arcs, Comput. Oper. Res. 40 (2013)
2975–2982.
[7] Derya Sever, Nico Dellaert, Tom Van Woensel Ton De Kok, Dynamic shortest path problems: hybrid routing policies considering network disruptions,
Comput. Oper. Res. 40 (2013) 2852–2863.
[8] Milad Nazarahari, Esmaeel Khanmirza, Samira Doostie, Multi-objective multi-robot path planning in continuous environment using an enhanced genetic
algorithm, Expert Syst. Appl. 115 (2019) 106–120.
[9] Morton Klein, A primal method for minimal cost flows with applications to the assignment and transportation problems, Manag. Sci. 14 (1967)
205–220.
[10] Andrew V. Goldberg, Robert E. Tarjan, Finding minimum-cost circulations by canceling negative cycles, J. ACM 36 (4) (1989) 873–886.
[11] Jack Edmonds, Richard M. Karp, Theoretical improvements in algorithmic efficiency for network flow problems, J. ACM 19 (2) (1972) 248–264.
[12] James B. Orlin, A polynomial time primal network simplex algorithm for minimum cost flows, Math. Program. 78 (1997) 109–129.
[13] I. Kuban Altınel, Necati Aras, Zeynep Şuvak, Z. Caner Taşkın, Minimum cost noncrossing flow problem on layered networks, Discrete Appl. Math. (2018).
[14] Tanka Nath Dhamala, A survey on models and algorithms for discrete evacuation planning network problems, J. Ind. Manag. Optim. 11 (1) (2015)
265–289.
[15] Wonjoon Choi, Horst W. Hamacher, Suleyman Tufekci, Modeling of building evacuation problems by network flows with side constraints, Eur. J. Oper.
Res. 35 (1) (1988) 98–110.
[16] Yue Liu, Gang-Len Chang, Ying Liu, Xiaorong Lai, Corridor-based emergency evacuation system for Washington, DC: system development and case
study, Transp. Res. Rec. 2041 (2008) 58–67.
[17] Yi Hong, Deying Li, Qiang Wu, Hua Xu, Dynamic route network planning problem for emergency evacuation in restricted space scenarios, J. Adv. Transp.
2018 (2018) 4295419, 13 pages.
[18] Yi Hong, Deying Li, Qiang Wu, Hua Xu, Priority-oriented route network planning for evacuation in constrained space scenarios, J. Optim. Theory Appl.
181 (2019) 279–297.
[19] Jihee Han, An efficient approach to 3D path planning, Inf. Sci. 478 (2019) 318–330.
[20] Mogens M. Graf Plessen, Partial field coverage based on two path planning patterns, Biosyst. Eng. 171 (2018) 16–29.
[21] Yi Hong, Jiandong Liu, Chuanwen Luo, Deying Li, Min-max-flow based algorithm for evacuation network planning in restricted spaces, in: COCOA 2018,
in: LNCS, vol. 11346, 2018, pp. 233–245.