Journal Pre-proof

Directed rooted forest based direction setting method: A step toward automated
dynamic exit signs

Botao Zhang, Jacqueline TY. Lo, Hongqiang Fang, Chuanzhi Xie, Tieqiao Tang,
Siuming Lo
PII: S2352-7102(24)00071-8
DOI: https://doi.org/10.1016/j.jobe.2024.108504
Reference: JOBE 108504

To appear in: Journal of Building Engineering

Received Date: 19 September 2023

Revised Date: 29 December 2023
Accepted Date: 10 January 2024

Please cite this article as: B. Zhang, J.T. Lo, H. Fang, C. Xie, T. Tang, S. Lo, Directed rooted forest
based direction setting method: A step toward automated dynamic exit signs, Journal of Building
Engineering (2024), doi: https://doi.org/10.1016/j.jobe.2024.108504.

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition
of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of
record. This version will undergo additional copyediting, typesetting and review before it is published
in its final form, but we are providing this version to give early visibility of the article. Please note that,
during the production process, errors may be discovered which could affect the content, and all legal
disclaimers that apply to the journal pertain.

© 2024 Published by Elsevier Ltd.

Title: Directed Rooted Forest Based Direction Setting Method: A Step Toward Automated
Dynamic Exit Signs

Botao Zhang, Ph. D. candidate

Department of Architecture and Civil Engineering, City University of Hong Kong
Email: [email protected]

Jacqueline T. Y. LO, Assistant Professor

Department of Civil and Environmental Engineering, Hong Kong Polytechnic University
Email: [email protected]

of
ro
Hongqiang Fang, Ph. D.

-p
Department of Architecture and Civil Engineering, City University of Hong Kong
Email: [email protected]
re
lP

Chuanzhi Xie, Ph. D. candidate

School of Transportation Science and Engineerging, Beihang University
na

Email: [email protected]
ur

Tieqiao Tang, Professor

Jo

School of Transportation Science and Engineerging, Beihang University

Email: [email protected]

Siuming Lo, Professor

Department of Architecture and Civil Engineering, City University of Hong Kong
Email: [email protected]
Botao Zhang1, Jacqueline TY Lo2*, Hongqiang Fang3, Chuanzhi Xie4, Tieqiao Tang5,
Siuming Lo6
1 Ph.D. Candidate, Department of Architecture and Civil Engineering, City University of Hong Kong, Hong
Kong, China; [email protected].
2 Assistant Professor, Department of Civil and Environmental Engineering, The Hong Kong Polytechnic
University, Hong Kong, China; [email protected].
3 Ph.D. Candidate, Department of Architecture and Civil Engineering, City University of Hong Kong, Hong
Kong, China; [email protected];
4 Ph.D. Candidate, School of Transportation Science and Engineering, Beihang University, Beijing, China;
[email protected].
5 Professor, School of Transportation Science and Engineering, Beihang University, Beijing, China;
[email protected].
6 Professor, Department of Architecture and Civil Engineering, City University of Hong Kong, Hong Kong,

of
China; [email protected]
*Correspondence: [email protected]; Tel. +852-2766-6033

ro
-p
re
lP
na
ur
Jo
1 Directed Rooted Forest Based Direction Setting Method: A
2 Step Toward Automated Dynamic Exit Signs
3

4 Abstract
5 Dynamic exit signs are generally considered to outperform static ones in guiding
6 evacuees toward exits in the changeable indoor environment during an emergency.
7 However, a crucial challenge lies in equipping them with a rapid and effective
8 adjustment method. This study proposes a directed rooted forest (DRF)-based planning

f
oo
9 method that can boost the efficiency of iterative optimization with heuristics from
10 simulation. In this method, the DRF structure is introduced to store direction

r
11 information to ensure the feasibility of the solution. Two algorithms, Branch Grafting
12
-p
and Leaf Grafting, are developed to enhance the global and local search capabilities.
re
13 Consequently, the optimizer composed of them converges fast by leveraging network
14 flow data obtained from the simulation in the optimization. Through testing, the method
lP

15 demonstrates a significant reduction in the number of simulation executions and

16 maintains a fast convergence rate when the problem size multiplies. In addition to the
na

17 engineering application potential offered by the accelerated simulation module, this

18 method holds promise in contributing to the development of a machine learning model
ur

19 for real-time guidance direction setting by rapidly generating optimized evacuation

Jo

20 guidance plans in a variety of scenarios for training purposes.

21 Keywords: Crowd Evacuation; Dynamic Exit Sign; Direction Setting; Directed

22 Rooted Forest (DRF)

23 1. Introduction
24 Every second counts in emergency evacuations. To reduce the time it takes for
25 people to escape from dangerous indoor spaces, scholars have attempted to optimize
26 evacuation from various aspects such as building layout, behavior management, and
27 strategy formulation [1]. Among those studies, one of the most practical and efficient
28 measures is to make better use of exit signs, which can shorten evacuees' decision-
29 making time and increase their moving speed [2,3]. Besides reasonable configuration
30 and strategic placement [4–6], they can be "smarter" with dynamically displayed
31 directions instead of static ones, which can only point to the nearest exit with a fixed

1
1 arrow regardless of the emergency condition[7]. Compared with the latter, the former,
2 which we will call automated dynamic exit sign (ADSE), can take into account the real-
3 time distribution of evacuees and the impact of hazards on the indoor passage capacity,
4 allocating the evacuees to different evacuation routes in a more balanced manner [8].
5 Multiple technologies and methods have been combined, from data acquisition to
6 generating an ad hoc evacuation plan consisting of all ADESs' directions. Thanks to
7 significant advancements in intelligent components such as sensors, UWB, WLAN, and
8 RFID[9,10], the acquisition and transmission of real-time data about evacuees,
9 buildings, and even hazards are no longer pain points [11,12]. Additionally, the progress
10 in pedestrian simulation models[13] has allowed researchers to consider the inherent
11 variation in human behavior and interactions with others, signage, and fire hazards [14].

f
oo
12 These advancements, along with improvements in hardware and software[15], have
13 facilitated the transition from mathematical planning based on evacuation networks [16]

r
14 to a double-layer planning approach [17], with an optimizer at the top layer and an
-p
15 evacuation simulation-based predictor at the sublayer [18]. By feeding the real-time
re
16 acquired information into the simulation model[19], more fine-grained evacuation
dynamics are captured, leading to a more reliable optimization process for evacuation
lP

17
18 guidance plans [20]. This planning framework serves as a solid foundation for the
ADES control algorithm [7].
na

19
20 However, the substantial computational time required to develop evacuation plans
ur

21 with these data hampers the introduction of "intelligence" into the exit signs. In the
22 aforementioned double-layer framework, as the prediction model estimates the impact
Jo

23 of the current plan on the evacuation process, the optimization method searches for a
24 new plan that may outperform the current one according to the estimation. Under this
25 feedback mechanism, repeated estimations are normally required before obtaining the
26 final plan. Due to the time-consuming nature of the prediction module, the number of
27 its executions, determined by the efficiency of the optimization method, can directly
28 impact the computational cost of planning[20]. Even if the prediction module can give
29 fast results, the cost becomes unacceptable if the optimization needs many attempts to
30 obtain a near-optimal plan since the input real-time data is already invalid by the time
31 the optimal plan is found [21].
32 In order to enhance the planning efficiency of double-layer planning based on
33 highly accurate evacuation prediction models and the applicability of automated
34 dynamic exit signs (ADES), several requirements must be fulfilled. Firstly, solutions
35 need to be stored in a more suitable data structure that allows for their transformation

2
1 into logical directions on exit signs. Secondly, the generation of invalid solutions should
2 be avoided by employing dedicated algorithms for searching for new solutions. Thirdly,
3 the convergence process should be accelerated by leveraging feedback information
4 from the prediction model. Lastly, the planning process should consider the crowd's
5 compliance rate with the exits to ensure its effectiveness, even in scenarios where partial
6 compliance is observed.
7 This study aligns with the current research trend of utilizing microsimulation
8 models to capture the interactions and dynamics of evacuations in various settings.
9 Building upon this, we address the requirements by making significant adaptations to
10 the optimizer to solve the simulation-based direction-setting problem. In contrast to
11 traditional methods, we propose a dedicated algorithm explicitly designed for solving

f
oo
12 the direction-setting problem. This algorithm accelerates the optimization process by
13 utilizing a more efficient solution storage structure and leveraging heuristic information.

r
14 To achieve this, we introduce the directed rooted forest (DRF) as a suitable encoding
-p
15 mechanism for the combination of directions on exit signs within given constraints. The
re
16 initial solution is obtained through the decomposition of the evacuation network, and a
reassignment algorithm is employed to search for the combinations that improve
lP

17
18 evacuation efficiency. Throughout this process, the multidimensional network flow data
obtained from the simulation serves two purposes: (1) evaluating the termination
na

19
20 condition and (2) determining the search direction. Furthermore, we incorporate a
ur

21 compliance rate parameter within the optimizer, which is dynamically adjusted based
22 on the mixed vector field extracted from the simulation. This incorporation enables the
Jo

23 planning method to effectively respond to partial compliance scenarios.

24 The structure of the paper is as follows: Section 2 provides a review of the solution
25 to the direction-setting problem and relevant evacuation planning problems. In Section
26 3, we delve into the detailed description of the DRF-based method and outline the
27 requirements for the simulation model. Section 4 presents the results obtained from
28 numerical tests and provides an analysis of the findings. Finally, in Section 5, we
29 conclude the paper and discuss potential avenues for future research.

30 2. Literature Review
31 As extensive research has revealed the limitation of fixed directional instructions
32 during an emergency evacuation, the dynamic exit signs are regarded as a crucial part
33 of an intelligent evacuation guidance system. While the design of acousto-optic
34 evacuation indicators with flash arrows or rotatable signs suspended from the ceiling

3
1 makes a wide range of directions displayable in practice, the critical problem becomes
2 how to build their 'mind' - direction setting algorithm [22].
3 In essence, the function of exit signs is to guide the evacuees to the limited available
4 evacuation routes. When representing routes with edges in the graph theory[23], the
5 direction setting problem (DSP) can be defined as a variant of the evacuation routing
6 problem (ERP) [24]. In ERP studies, the multi-source and multi-sink evacuation
7 scenario is transferred into a time-varying network, and the algorithms for the quickest
8 flow problem are consequently introduced to solve such problems [25–27]. In
9 comparison, the DSP, by nature, is hop-by-hop routing where a similar network can be
10 adopted but with two additional constraints: (1) the nodes represent the positions of
11 signs; (2) the outflow from the node cannot be precisely controlled [28]. While the first

f
oo
12 constraint slightly reshapes the network, the second constraint completely transforms
13 the problem into a hierarchically solved shortest-path problem. Under these constraints,

r
14 the initial approach was to generate all valid sign direction setting scenarios and
-p
15 evaluate them through evacuation time calculations[29]. However, the feasibility of this
re
16 method diminishes as the network size increases. To address this problem, Cho et al.
[30] compared the applicability and computational complexity of four shortest-path
lP

17
18 algorithms and pioneered the development of the direction-setting algorithm based on
Dijkstra's algorithm. At the implementation level, Kim et al. [31] proposed a prototype
na

19
20 of the intelligent exit sign system, including a modified algorithm and a sever-
ur

21 independent hardware configuration. To ensure the system's stability under

22 disadvantageous operating conditions, they develop decentralized signs that gather data
Jo

23 from neighboring nodes and employ the Breadth First Searching (BFS)-like algorithm
24 to find the shortest path. Conversely, Fu et al. [32] favored a server-dependent sign
25 system that captures the complexity of the building layout through a weighted
26 navigation network generated by the BIM model. In addition, in their study, BFS (for
27 eliminating unsafe paths) and Dijkstra's algorithm (for finding the shortest path) are
28 combined to improve the method's efficiency. Building upon the same concept,
29 Mirahadi and McCabe [33] utilized the integrated modules mentioned above to
30 determine the direction of ADSEs and presented a case study showcasing the
31 improvement in evacuation safety. In the above studies, it is noticeable that the
32 movement of evacuees is significantly simplified in the increasingly precise network.
33 In their hypothesis, evacuees escape in an ideal pattern, transferring between nodes
34 along the edge at a fixed/ density-dependent speed [24]. While this assumption

4
1 substantially reduces the computational cost, it may lead to an incorrect estimation of
2 the evacuation process, especially in situations with high crowd density [21,34].
3 Placing importance on the specific behaviors of evacuees who exactly receive the
4 guidance, another research idea, known as simulation-based planning, is gradually
5 emerging [35]. In a self-organized evacuation simulation, evacuees' movements are
6 determined by the tactical layer, where they perceive the "big picture" and select the
7 path at the operational layer, where they adjust their position according to the actual
8 surroundings [36,37]. Consequently, from a planning standpoint, the ERP that considers
9 evacuee dynamics is constructed in a similar structure. In this type of problem,
10 optimization methods operate at the tactical layer to suggest the optimal paths to
11 evacuees, and the dynamics of evacuees under recommendations are reproduced with

f
oo
12 the pedestrian simulation models at the operational layer.
13 To granularly depict how individuals interact with others and their environment

r
14 during evacuation, various pedestrian simulation models have been integrated into the
-p
15 above framework [8,38–42]. Notably, the Social Force Model (SFM) and the Cellular
re
16 Automata (CA) model have gained significant popularity as they represent continuous
and discrete models, respectively [15,20,21]. These models provide a comprehensive
lP

17
18 representation of the interactions between evacuees and their surroundings, including
other evacuees, obstacles, and signage. In the SFM model, pedestrian agents move
na

19
20 toward their targets driven by the desire to reach their destination. However, their
ur

21 directions are influenced by repulsive forces generated by surrounding obstacles and

22 other pedestrians. This model has garnered substantial attention due to its intuitive
Jo

23 physical interpretation. Extensive controlled experiments and field observations have

24 been conducted to calibrate model parameters and devise new rules to address inherent
25 limitations such as jumping, overlapping, and rebounding [43]. While the SFM
26 represents a prominent example of continuous models, the CA model stands as the most
27 widely used discrete model in pedestrian simulation. In the CA model, pedestrians are
28 influenced by both static fields, which consider obstructions in the environment, and
29 dynamic fields, representing the "chemical influence" derived from the behavior of
30 surrounding pedestrians. Pedestrians are then moved within the lattice points based on
31 a transition matrix. The CA model has advantages in terms of its scalable rules and its
32 efficiency in conducting large-scale simulations [44].
33 Optimization methods have been developed and refined to effectively utilize
34 micromodels in evacuation planning and address a range of challenges. The origins of
35 these developments can be traced back to the optimization-simulation-optimization

5
1 structure introduced by Park et al. [39] In their optimizer, the updated path weights
2 obtained from micro-simulation enable Dijkstra's algorithm to continually search for
3 improved recommendation paths according to the evacuation network (EN) at the top
4 layer. Chu et al.[41] further elucidated the two-layer structure in the solution of the ERP
5 by combining the Markov decision-based compliance model proposed by Luh et al [45]
6 with the microsimulation model. In addition to the progressively refined lower-level
7 simulation models, upper-level optimization models incorporate meta-heuristic
8 algorithms (e.g., simulated annealing algorithm [46], ant colony algorithm[47], tabu
9 search[48], and artificial immune system[49]). Despite their stochastic nature, meta-
10 heuristics can reduce the computational complexity in solving large-scale problems and
11 act as black boxes irrespective of the problem structure[49]. These characteristics make

f
oo
12 them well-suited for solving the exits assignment problem (EAP), where the geometric
13 network graph does not impose strict constraints on the form of the solution. Stochastic

r
14 search in meta-heuristic algorithms enables iterative optimization with high
-p
15 accuracy[50]. When applied to solve DSP with more aforementioned nonlinear
re
16 constraints, the meta-heuristic algorithms provide researchers with a wider range of
evaluation criteria. An example of this is the work of Li et al., who employed the
lP

17
18 genetic algorithm (GA) in searching for the directed graph, which can improve
evacuation efficiency by reducing congestion and enhancing the level of service (LOS)
na

19
20 in large-scale facilities [51]. Furthermore, they allow researchers to transcend the
ur

21 conventional notion of orienting signs according to the paths in the EN. A case in point
22 is Li et al.'s work, in which both the layout and orientation of exit signs are
Jo

23 synchronously optimized with the GA[52]. As outlined, a typical optimization

24 framework for DSP contains the meta-heuristic algorithm in the top layer, and an
25 evacuation simulation model in the bottom layer is established [8,42]. The simulation
26 prediction results, which can reflect the evacuee dynamics under the influence of signs,
27 are fed back to the optimization layer as heuristic information to guarantee convergence
28 towards an approximate optimal solution within a finite time.

29 Table 1. Comparison between ERP, EAP, and DSP.

ERP EAP DSP

6
 Illustration

Feature Network-dependent Network-independent Network-dependent

Greedy Algorithms
Greedy Algorithms
Meta-heuristic /
/
Algorithms Meta-heuristic
Solution Greedy Algorithms
+ Algorithms
+

f
Crowd Simulation +

oo
Crowd Simulation
Crowd Simulation

r
1
2
-p
The comparison between ERP, EAP, and DSP is shown in Table 1. The solution of
DSP is inspired by the exploration of ERP and EAP. While greedy algorithms can find
re
3
4 the optimal solution quickly in small-scale networks, two-layer optimization more
lP

5 accurately replicates evacuee dynamics and obtains satisfactory solutions within a

6 linearly increasing time cost[16]. However, the network-dependent nature of the
na

7 problem prevents meta-heuristic algorithms from fully exploiting their advantages.

8 Since the simulation only provides less heuristic information, the search direction is
ur

9 predominantly random. Additionally, the random search yields a relatively low

Jo

10 proportion of valid solutions due to the constraints imposed by the network structure.
11 In conclusion, a dedicated algorithm for solving the DSP is still urgently needed.

12 3. Methodology
13 The main objective of the method proposed in this study is to establish the direction
14 information on exit signs by utilizing collected data pertaining to the building, the
15 crowd, and the disaster situation. During this process, the aforementioned input
16 information is utilized to construct an evacuation network, forming the foundation for
17 initially obtaining a valid direction-setting plan using the Directed Rooted Forest (DRF)
18 approach. Next, the performance of the DRF-based plan in the current scenario is
19 evaluated through an evacuation simulation. Throughout the simulation, the optimizer
20 continuously enhances the structure of the DRF based on the intermediate and final
21 outputs until the termination condition is met. Lastly, the optimized DRF is transformed
22 into the direction information displayed on the Automated Dynamic Exit Signs (ADES).

7
1 The technical roadmap of this method is depicted in Fig. 1.
2 The technical roadmap consists of several essential steps, which are elaborated in
3 the following sections. Subsection 3.1 outlines the process of creating a sign-oriented
4 evacuation network and the corresponding constraints that must be met to establish the
5 direction information on the signs. Subsection 3.2 explains the methodology for setting
6 the direction of nodes within the network to optimize efficiency during the evacuation
7 process. Subsection 3.3 discusses the required simulation outputs and the algorithms
8 employed to utilize these feedbacks as references for efficient and purposeful
9 optimization of the current direction-setting plan. Lastly, subsection 3.4 provides a
10 supplementary description of acquiring the necessary optimization parameters based on
11 the mechanisms integrated within the simulation model. The meanings of the symbols

f
oo
12 used therein, in addition to the elaboration in this chapter, can be found in Appendix
13 Table 1A.

r
14
-p
re
lP
na
ur
Jo

15
16 Fig. 1. The technical roadmap of the sign setting optimization method

17 3.1 Evacuation network construction

8
1 When the internal layout in the building information and the hazard information
2 are available in real-time, the location of the available signs is usually taken as the node
3 in the EN construction for DSP. However, this method alters the construction of EN by
4 modifying the space partitioning process. Each subarea is intended to accurately
5 encompass the influence range of its corresponding sign on the crowd for two reasons:
6 (1) the crowd dynamics generated by the simulation within each subarea will be
7 aggregated, and (2) all evacuees within that subarea are considered as a synchronously
8 assigned unit in the optimization. Therefore, we refer to the visibility catchment area
9 (VCA) to determine the area in which evacuees can visually receive information on the
10 sign and be directly affected by it at the node [53]. In addition, evacuees whose actions
11 are affected by the signs' instructions are not limited to those who visually perceive it.

f
oo
12 Due to the memory effect and following behavior, the actual affecting range of signs
13 could be expanded by taking the centroid of all VCAs in each room as seeds for the

r
14 Voronoi partition of the space (as depicted in Fig. 2). On this basis, the centroids of
-p
15 subareas are regarded as the internal nodes in the EN and the midpoints of the exit
re
16 sections as the exit nodes. Let 𝑁 ∈ {n1 , n2 , ⋯ , n𝑚 } represent the set of all nodes in the
EN and 𝐸 ∈ {e1 , e2 , ⋯ , e𝑛 } the set of exit nodes. Because of the mapping from 𝐸 to 𝑁,
lP

17
18 the subscript of e𝑘 ’s image in 𝑁 can be given by 𝑖𝑛𝑑(e𝑘 ). For each node n𝑖 in 𝑁, its
corresponding subarea is denoted by s𝑖 (it should be noted that the subarea
na

19
20 corresponding to the exit node contains only one straight line, i.e. the projection of the
ur

21 exit section in the plane). If two subareas n𝑖 and n𝑗 share an unobstructed boundary b𝑖𝑗 ,
22 the corresponding nodes become neighbors and are added to each other into their
Jo

23 adjacency lists 𝐴𝑖 and 𝐴𝑗 . Subsequently, edges are generated between all the
24 neighboring nodes whose lengths are weighted values in the adjacency matrix 𝐷. The
25 DSP's solution is converting these edges into valid directed edges, so that the direction
26 on sign at the tail will be set to point to the midpoint of the boundary between the nodes
27 at the two ends. The criteria of validity in this setting method and the resulting
28 constraints are described herein below.

9
 f
oo
1
2 Fig. 2. Signs-oriented space partitioning and network construction.

r
3 -p
According to the assumptions of DSP, evacuees within those divided sub-areas will
4 choose their direction of escape by referring to the direction indicated by the signs.
re
5 Therefore, it is crucial that the signs' directions guide evacuees from any initial position
lP

6 towards an accessible exit. To effectively meet this requirement, Fu and Liu have
7 established three crucial criteria for evaluating the effectiveness of the exit sign system
na

8 [54]:
9 (1) Continuity: Clear direction-indicator should be given at each key position.
ur

10 (2) Direction consistency: No conflicting directions are on adjacent signs.

(3) Guiding correctness: They help evacuees escape as quickly as possible.
Jo

11
12 The system that fails to meet these requirements will lead to chaotic evacuation, as
13 depicted by the crowd's escape trajectory in Fig. 3. While the discontinuous or
14 inconsistent signage systems may cause evacuees to wander, the incorrect signage
15 systems can lead to highly ineffective detours. It is important to clarify that an
16 ineffective detour here refers to a detour that does not reduce evacuation time. If
17 following detour routes helps to avoid congestion and reduce evacuation time, setting
18 up signs according to these routes is acceptable. However, under no circumstances
19 should a circuitous detour that would lead evacuees in circles be created.

10
 f
r oo
1 -p
2 Fig. 3. Schematic diagram of invalid sign settings: (a) discontinuous; (b) inconsistent;
re
3 (c) incorrect, and their resulting evacuees’ trajectory: (d), (e), (f).
lP

4 To satisfy the above constraints, the concept of directed rooted forests (DRF), a
5 disjoint union of directed rooted trees (DRTs), is introduced. Treating each exit node as
na

6 the root node, the decomposition of the evacuation network into a DRF of size |𝐸| can
7 avoid the appearance of loop guidance that does not satisfy the guiding correctness. In
ur

8 addition, the DRF used here is an in-forest with an anti-branching structure (i.e., all
Jo

9 edges within the rooted trees point towards their respective roots). The choice of such
10 a forest structure ensures that both continuity and direction consistency can be met. As
11 described, the solution of DSP is translated into how to decompose the EN into DRF
12 (as shown in Fig. 4.) that maximizes evacuation efficiency.

11
 f
r oo
-p
re
lP
na
ur

1
2 Fig. 4. Schematic diagram of the EN-to-DRF decomposition.
Jo

3 To obtain a spanning forest structure that is easier to maintain data-wise, the Prim
4 algorithm is more suitable than the Kruskal algorithm[55]. Consequently, the adjacency
5 list should be constructed beforehand to reduce the time complexity of the algorithm.
6 Because of the uncertain dynamics of evacuees, assigning nodes to the trees in the forest
7 can hardly be accomplished at one stroke. The nodes stored in the forest need to be
8 continuously relocated until the metrics output from the simulation satisfies the
9 termination condition (shown in Fig. 1).
10 The whole direction setting method takes the EN as input and the optimal DRF
11 with directions indicated on dynamic signs at each node as output. Such method consists
12 of an assignment method for DRF initialization and a reassignment algorithm for DRF
13 optimization. Moreover, it can be seamlessly integrated with different simulation
14 models.

15 3.2 DRF Initialization

12
1 As mentioned, the DRF is a disjoint union of DRTs characterized by the anti-
2 branching structure. So, the basic initialization requirement is to assign nodes to DRTs
3 in a logical order. In addition to grouping them into the trees 𝑇1 , 𝑇2 , ⋯ 𝑇|𝐸| which
4 embody the node-to-tree affiliation information, the hierarchical parent-child
5 relationships of each node on the tree are also recorded with the parent-child hierarchy.
6 Given that all nodes in anti-branching, except the root nodes with the outdegree of 1,
7 the only one parent node of n𝑖 , 𝑖 ∈ {1,2, ⋯ , |𝑁|} is marked as n𝑃𝑖 .
8 Besides the basic requirements, another goal of initialization is to obtain a solution
9 as close to the optimal solution as possible. Therefore, basic information about network
10 flow should be collected from the EN before the assignment. As part of the preparation,
11 the numbers of evacuees initially located in 𝑠𝑖 , 𝑖 ∈ {1,2, ⋯ , |𝑁|} are counted as 𝑥𝑖 .

f
oo
12 Using the Dijkstra's algorithm with the adjacency matrix 𝐷 as the input, 𝑑𝑖,𝑘 , 𝑖 ∈
13 {1,2, ⋯ , |𝑁|}, 𝑘 ∈ {1,2, ⋯ , |𝐸|} which denote the geodesic distances from n𝑖 to e𝑘 can

r
14 be calculated. Subsequently, the movement characteristics of pedestrians during
-p
15 evacuation need to be extracted from the fundamental diagram[56], including average
re
16 velocity 𝑣 and maximum flow rate per unit 𝑞𝑚𝑎𝑥 . By considering the width of the exit,
the evacuation capacity of each exit 𝑟𝑘 , 𝑘 ∈ {1,2, ⋯ , |𝐸|} can be obtained.
lP

17
18 Utilizing the aforementioned data, we can allocate all nodes in the DRF with an
assignment algorithm that relies on the evacuation time prediction results estimated by
na

19
20 a simplified linear model for the narrow door problem. In this model, the time it takes
ur

21 for all evacuees in a node to evacuate the building consists of two linear components:
22 the time to reach the exit and the time to queue at the exit in the presence of congestion.
Jo

23 When independently considering that evacuees in n𝑖 are assigned to leave the building
24 from e𝑘 , 𝑡𝑘 (n𝑖 ), which represents the evacuation time for the last evacuee in the node
25 to exit the building, can be determined by the following equation [26].
𝑑𝑖.𝑘 𝑥𝑖
26 𝑡𝑘 (n𝑖 ) = + (1)
𝑣 𝑟𝑘
27 On this basis, when evacuees in another node n𝑗 are also assigned to the same exit,
28 the evacuation time can be updated according to the superposition of the linear model,
29 as illustrated in the following equation.
𝑑𝑗,𝑘 𝑥𝑗
max {𝑡𝑘 (n𝑖 ), }+ , if 𝑑𝑖,𝑘 ≤ 𝑑𝑗,𝑘
𝑣 𝑟𝑘
30 𝑡𝑘 (n𝑖 , n𝑗 ) = (2)
𝑑𝑖,𝑘 𝑥𝑖
max {𝑡𝑘 (n𝑗 ), }+ , if 𝑑𝑖,𝑘 > 𝑑𝑗,𝑘
{ 𝑣 𝑟𝑘

31 According to the above model, unassigned nodes can be sequentially assigned to

32 the DRTs with the shortest current evacuation time according to the distance to the

13
1 corresponding exit. The algorithm for DRF Assignment is therefore proposed and
2 shown in Algorithm 1.
3
Algorithm 1 DRF Assignment
Input: 𝐴𝑖 for n𝑖 ∈ 𝑁 (Adjacency List)
& 𝐷 (Weighted adjacency matrix)
& 𝑥𝑖 for n𝑖 ∈ 𝑁 (Evacuee Distribution)
Output: 𝑇𝑘 for e𝑘 ∈E (Node-to-Tree Affiliation)
& n𝑃𝑖 for n𝑖 ∈ 𝑁 (Parent-Child Hierarchy)
1: Initialization 𝑇𝑘 ← {e𝑖 } for e𝑘 ∈E
2: Create temporary list 𝐿∗ ← {0} × |𝐸|

f
3: Create temporary lists 𝐶𝑘∗ ← 𝑙𝑖𝑠𝑡(𝐴𝑖𝑛𝑑(e𝑘) ) for e𝑘 ∈E

oo
4: Create temporary lists 𝑃𝑘∗ ← {e𝑘 } × |𝐶𝑘 | for e𝑘 ∈E
5: Create temporary list 𝑈 ∗ ← 𝑙𝑖𝑠𝑡(𝑁 − 𝐸)

r
6: while 𝑈 ∗ ≠ ∅ do
7: α ← argmin{𝐿∗ [𝑘]}
-p
re
𝑘∈{1,2,⋯|𝐸|}
8: 𝛽 ← argmin {𝑑𝑖𝑠(𝐶𝛼∗ [𝑖], e𝛼 )}
𝑖∈{1,2,⋯|𝐶𝛼∗ |}
lP

9: 𝛾← 𝑖𝑛𝑑(𝐶𝛼∗ [𝛽])
∗
10: if n𝛾 ⊆ 𝑈 then:
na

11: 𝑈 ∗ ← 𝑈 ∗ − {n𝛾 }
12: 𝐶𝛼∗ ← 𝐶𝛼∗ + 𝑙𝑖𝑠𝑡(𝐴𝛾 )
ur

13: 𝑃𝛼∗ ← 𝑃𝛼∗ + {n𝛾 } × |𝐴𝛾 |

Jo

14: 𝑇𝛼 ← 𝑇𝛼 + {n𝛾 }
15: n𝑃𝛾 ← 𝑃𝛼∗ [𝛽]
16: 𝐿∗ [𝛼] ← max{𝐿∗ [𝛼], 𝑑𝛾,𝛼 ⁄𝑣 } + 𝑥𝛾 ⁄𝑟𝛼
17: end if
18: delete 𝐶𝛼∗ [𝛽]
19: delete 𝑃𝛼∗ [𝛽]
20: end while
4
5 Among the two outputs generated by the DRF Assignment, obtaining the node-to-
6 tree affiliation through the parent-child hierarchy is feasible and advantageous due to
7 its higher information entropy resulting from its explicit structure. Nevertheless, the
8 affiliation, which serves as an intermediary output of the algorithm, is preserved due to
9 its ability to reduce the search's computational cost in the subsequent optimization. It is
10 represented by a partition of 𝑁 that consists of a group of mutually disjoint subsets

14
1 𝑇1 , 𝑇2 ⋯ 𝑇|𝐸| governed by ⋃𝑖∈{1,2,⋯|𝐸|} 𝑇𝑖 = 𝑁 and 𝑇𝑖 ∩ 𝑇𝑗 = ∅, 𝑖 ≠ 𝑗. Each node in the
2 EN is exclusively assigned to a tree set, which can be traced with the function 𝑡𝑟𝑒𝑒().
3 Another output of the assignment is the parent-child hierarchy, which provides
4 valuable insights into the positional relationships among nodes within the tree. In the
5 first place, the hierarchy is recorded with a parental presentation that facilitates the
6 process of conducting depth-first traversal. Equivalently, it can be transformed into a
7 child representation suitable for conducting breadth-first traversal. In such a structure,
8 all the child nodes of n𝑖 should be put into set 𝐶𝑖 = {n𝑗 ∈ 𝑡𝑟(n𝑖 )| n𝑗𝑃 = n𝑖 } . If this set
9 is empty, it means that the in-degree of the node is 0 and the node is therefore considered
10 as a leaf node of 𝑡𝑟𝑒𝑒(n𝑖 ) and is included in the set 𝐿𝑘 . Among all nodes in the 𝑇𝑘 ,
11 excluding the set of leaf nodes 𝐿𝑘 and the root node e𝑘 , the remaining nodes in between
12 are referred to as branch nodes, and their set is denoted by 𝐵𝑘 = 𝑇𝑘 ∖ (𝐿𝑘 ∪ {e𝑘 }).

f
oo
13 Algorithm 1 aids in the initial transformation of the EN into a compliant DRF by
14 incorporating a prediction model is introduced to achieve a rough balance of the

r
15 evacuation time at each exit. Still, its prediction accuracy is inferior to that of the micro-
16
-p
simulation model. The interactions between evacuees, between evacuees and obstacles,
17 and between evacuees and instructional information are disregarded due to the fixed
re
18 values of speed, distance, and number of evacuees following instructions. Therefore, an
lP

19 algorithm that can improve DRF based on more detailed microsimulation results is
20 needed.
na

21 3.3 DRF Optimization

22 With the initial DRF, the direction on the dynamic signs can be set in align with the
ur

23 parent-child hierarchy, specifically from children to parents. The performance of the

Jo

24 current direction-setting plan can be evaluated in the simulation model, and related data
25 will be fed into the optimization module. Apart from the evacuation times of each exit,
26 the heuristic information required for subsequent optimization includes the number of
27 evacuees flowing out from each node in the EN. To be specific, the cumulative outflows
28 from each internal node throughout the evacuation process are measured collectively,
29 and this is referred to as the undertaking volume. In contrast, exit nodes' outflows are
30 measured in time periods to assess each tree's root evacuation capacity. For
31 differentiation, the tilde “∼” is employed to label the estimated value from the virtual
32 environment, including simulation output and derived data. In this way, the undertaking
33 volume 𝑦̃𝑖 of the node n𝑖 can be counted by summing up outflow at each boundary
34 denoted by 𝑦̃𝑖𝑗 for n𝑗 ∈ 𝐴𝑖 . Moreover, the outflows from the k-th exit measured in time
35 slots are recorded in a time series 𝑅̃𝑘 = {𝑟̃1𝑘 , 𝑟̃2𝑘 , ⋯ , 𝑟̃𝑚𝑘 } where the total evacuation time
36 of the k-th exit node 𝑡̃𝑘 is divided into 𝑚 segments of equal duration. They are governed
37 by the following equation.

15
 𝑛
1 ∑ 𝑟̃𝑖𝑘 ∙ 𝑡̃𝑘 ⁄𝑚 = 𝑦̃𝑖𝑛𝑑(𝑒𝑘) (3)
𝑖=1
2 As evacuees are required to exit the facility via the exit nodes, the evacuation
3 time of the root node, the longest one in a tree, can reflect the utilization of this exit
4 in the evacuation under the current setting. Similar to how the capacity of Liebig's
5 barrel is determined by the shortest stave, the evacuation time of the entire building
6 is determined by the longest evacuation time of all root nodes in the DRF, considering
7 the last evacuee's departure from the hazardous indoor environment. This to-be-
8 reduced indicator is called the evacuation time of the DRF, represented by 𝑡̃ F and
9 given by
10 𝑡̃ F = max {𝑡̃𝑘 } (4)

f
𝑘∈{1,2,⋯|𝐸|}

oo
11 According to the “Uniformity Principle” [57], the reduction of 𝑡̃ F can be achieved by
12 making trade-off between 𝑡̃𝑘 , 𝑘 ∈ {1,2, ⋯ |𝐸|} corresponding to the estimated

r
13
14
-p
evacuation time of evacuees through different exits. To reduce the deviation of the
maximum value, we introduce the sum of squared deviations from mean (SSDM).
re
|𝐸| 2
15 ̃ =∑
𝑆𝑆 (𝑡̃𝑘 − 𝑡̃̅) (5)
lP

𝑘=1
16 where 𝑡̃̅ denotes the average evacuation time of all root nodes in the DRF, and it can be
17 expanded into
na

|𝐸| |𝐸| 2
1
18 ̃=
𝑆𝑆 ∑ (∑ (𝑡̃𝑘 − 𝑡𝑗̃ )) (6)
|𝐸| 𝑘=1
ur

𝑗=1

19 This method reduces SSDM by narrowing the largest difference between

Jo

20 evacuation times of root nodes, which is achieved by altering the affiliations of the
21 nodes. Specifically, certain nodes will be selected and “pruned” from the tree with a
22 longer evacuation time, and then they will be reassigned as new children of a node on
23 the tree with a shorter evacuation time. If the selected node is a branch node, to ensure
24 connection continuity, all its descendants will also retain the previous parent-child
25 hierarchy. In such case, a chain of nodes is reassigned simultaneously, which can be
26 further classified as Branch Grafting. By contrast, when the selected node is a leaf node
27 on the original tree, the grafting process is called Leaf Grafting. Although there are no
28 restrictions on the type of node, when selecting, two conditions must be met:
29 (1) The adjacent set of the selected node should contain a node on another tree with a
30 lower evacuation time (i.e., ∃n𝑢 ∈ 𝐴𝑖 ∩ 𝑇ℎ , 𝑡̃ℎ < 𝑡̃𝑘 , n𝑖 ∈ 𝑇𝑘 );
31 (2) The relocation should not result in the evacuation time of the target tree to
32 significantly exceeding the evacuation time of the original tree, thus leading to an

16
 ′ ′
1 increased difference between the two(i.e., 𝑡̃ℎ − 𝑡̃𝑘 < 𝑡̃𝑘 − 𝑡̃ℎ ).
2 In determining condition (1), each node needs to calculate the maximum difference
3 ∆𝑡̃𝑖 based on its own tree and its neighboring nodes.
4 ∆𝑡̃𝑖 = max {∆𝑡̃𝑖𝑗 } = max {𝑡̃𝑘 − 𝑡̃ℎ |n𝑖 ∈ 𝑇𝑘 , n𝑗 ∈ 𝑇ℎ } (7)
n𝑗 ∈𝐴𝑖 n𝑗 ∈𝐴𝑖

5 Only when ∆𝑡̃𝑖 > 0, it indicates that condition (1) is satisfied and thus allows for the
6 evaluation of condition (2). For either type of Grafting, the absence or invariance of
7 selected node’s descendants guarantees that its undertaking volume remains unchanged.
8 Hence, the evacuation time of the original tree (e.g., the k-th tree) and that of the target
9 tree (e.g., the h-th tree) after Grafting can be evaluated according to the following
10 formulas:

f
oo
′
11 𝑡̃𝑘 = 𝑡̃𝑘 − 𝜀𝑖𝑗 ∙ 𝑦̃𝑖 ⁄𝐹(𝑅̃𝑘 ) (8)

′
𝑡̃ℎ = 𝑡̃ℎ + 𝜀𝑖𝑗 ∙ 𝑦̃𝑖 ⁄𝐹(𝑅̃ℎ )

r
12 (9)

13
-p
Where 𝜀𝑖𝑗 represents the proportion of 𝑦̃𝑖 that are transferred to the target tree when n𝑗 ,
re
14 a node in that tree, becomes n𝑖 ’s parent node. These coefficients should be dedicated
15 set before reassignment and continuously adjusted after each round of simulation. 𝐹(𝑅̃)
lP

16 is a function for extracting the average outflow rate at each exit under the congestion
17 with the time series 𝑅̃. In this function, moving average of order 𝜎 is utilized as per
na

18 equation (8). The reason for adopting this rate is that the transferred nodes are located
19 at the leaf end of the target tree (i.e., at a location far from the exit), which results in a
ur

20 lengthening of the congestion phase.

Jo

𝑘
1 ⌊𝜎/2⌋ 𝑚 ∙ 𝑟̃𝑡+𝑗
21 ̃
𝐹(𝑅𝑘 ) = max { ∑ } (10)
𝜎 𝑗=−⌊𝜎/2⌋ 𝑡̃𝑘

22 Based on the above estimates, condition (2) can be expressed as a requirement for
23 the undertaking volume of the selected node as following:
2(𝑡̃𝑘 − 𝑡̃ℎ )
24 𝑦̃𝑖 < (11)
𝜀𝑖𝑗 ∙ [1⁄𝐹(𝑅̃𝑘 ) + 1⁄𝐹(𝑅̃ℎ )]
25 The right-hand side of the above inequality is defined as 𝑦̃𝑖𝑚𝑎𝑥 . It can be found that
𝑜𝑝𝑡
26 there exists an optimal value of undertaking volume 𝑦̃𝑖 = 𝑦̃𝑖𝑚𝑎𝑥 ⁄2 that can make
′ ′ 𝑜𝑝𝑡
27 𝑡̃𝑘 = 𝑡̃ℎ . So, the value of proximity 𝜂𝑖 of 𝑦̃𝑖 to 𝑦̃𝑖 is introduced to further screen
28 nodes in Branch Grafting.
𝑜𝑝𝑡
|𝑦̃𝑖 − 𝑦̃𝑖 |
29 𝜂̃𝑖 = 1 − 𝑜𝑝𝑡 (12)
𝑦̃𝑖
30 Based on the above discussion, we have developed two sub-algorithms, namely

17
1 Branch Grafting and Leaf Grafting, for DRF reassignment. The details of these sub-
2 algorithms are listed in Algorithm 2.1 and Algorithm 2.2.
3
Algorithm 2.1 Branch Grafting
Input: 𝑇𝑖 for e𝑖 ∈E (Node-to-Tree Affiliation)
& n𝑃𝑖 for n𝑖 ∈ 𝑁 (Parent-Child Hierarchy)
& 𝑡̃𝑘 , 𝑅̃𝑘 for 𝑘 = {1,2, ⋯ , |𝐸|} & 𝑦̃𝑖 for n𝑖 ∈ 𝑁 (Simulation Result)
Output: 𝑇𝑖 for e𝑖 ∈E (Node-to-Tree Affiliation after Branch Grafting)
& n𝑃𝑖 for n𝑖 ∈ 𝑁 (Parent-Child Hierarchy after Branch Grafting)
& 𝑡̃𝑘′ for 𝑘 = {1,2, ⋯ , |𝐸|} (Predicted evacuation time after Branch Grafting)
& 𝜊 𝐵 (Observer for Branch Grafting execution)

f
1: Initialization 𝜊 𝐵 ← 𝐹𝑎𝑙𝑠𝑒

oo
2: Create temporary list 𝑆 ∗ ← ⋃𝑖∈{1,2,⋯|𝐸|} 𝐵𝑖
3: Calculate ∆𝑡̃𝑖 for n𝑖 ∈N with Equation (5)

r
-p
4: Calculate 𝜂̃𝑖 for n𝑖 ∈N with Equation (10)
5: 𝑚 ← max∗ ∆𝑡̃𝑖
re
n𝑖 ∈𝑆
6: for all n𝑖 in 𝑆 ∗ do
lP

7: if ∆𝑡̃𝑖 < 𝑚 or Equation (11) does not hold then:

8: 𝑆 ∗ ← 𝑆 ∗ − {n𝑖 }
9: end if
na

10: end for

11: if 𝑆 ∗ ≠ ∅ do
ur

12: 𝜊 𝐵 ← 𝑇𝑟𝑢𝑒
Jo

13: 𝛼 ← argmax{𝜂̃𝑖 }
𝑖 for n𝑖 ∈𝑆 ∗

14: 𝛽 ← argmax {∆𝑡̃𝑖𝑗 }

𝑗 for n𝑗 ∈𝐴𝑖

15: n𝑃𝛼← n𝛽
16: Get {𝑡̃𝑘′ } for 𝑘 = {1,2, ⋯ , |𝐸|} with Equation (8, 9)
17: Create temporary list 𝐺 ∗ ← {n𝛼 }
18: while 𝐺 ∗ ≠ ∅ do
19: n𝛾 ← 𝐺 ∗ [0]
20: 𝑡𝑟𝑒𝑒(n𝛼 ) ← 𝑡𝑟𝑒𝑒(n𝛼 ) − {n𝛾 }
21: 𝑡𝑟𝑒𝑒(n𝛽 ) ← 𝑡𝑟𝑒𝑒(n𝛽 ) + {n𝛾 }
22: for all n𝑖 in 𝐶𝛾 do
23: 𝐺 ∗ ← 𝐺 ∗ + {n𝑖 }
24: end for
25: 𝐺 ∗ ← 𝐺 ∗ − 𝐺 ∗ [0]

18
 26: end while
27: end if
1
Algorithm 2.2 Leaf Grafting
Input: 𝑇𝑖 for e𝑖 ∈E (Node-to-Tree Affiliation)
& n𝑃𝑖 for n𝑖 ∈ 𝑁 (Parent-Child Hierarchy)
& 𝑅̃𝑘 for 𝑇𝑘 ∈ 𝐹 and 𝑦̃𝑖 for n𝑖 ∈ 𝑁 (Simulation Result)
& 𝑡̃𝑘′ for 𝑘 = {1,2, ⋯ , |𝐸|} (Predicted evacuation time after Branch Grafting)
Output: 𝑇𝑖 for e𝑖 ∈E (Node-to-Tree Affiliation after Leaf Grafting)
& n𝑃𝑖 for n𝑖 ∈ 𝑁 (Parent-Child Hierarchy after Leaf Grafting)
& 𝑡̃𝑘′ for 𝑘 = {1,2, ⋯ , |𝐸|} (Predicted evacuation time after Leaf Grafting)

f
& 𝜊 𝐵 (Observer for Leaf Grafting execution)

oo
1: Initialization 𝜊 𝐿 ← 𝐹𝑎𝑙𝑠𝑒
1: Create temporary list 𝑆 ∗ ← ⋃𝑖∈{1,2,⋯|𝐸|} 𝐿𝑖

r
-p
2: Calculate ∆𝑡̃𝑖 for n𝑖 ∈N with Equation (5)
3: Calculate 𝜂̃𝑖 for n𝑖 ∈N with Equation (10)
re
4: 𝑚 ← max∗ ∆𝑡̃𝑖
n𝑖 ∈𝑆
lP

5: for all n𝑖 in 𝑆 ∗ do
6: if ∆𝑡̃𝑖 < 𝑚 or Equation (11) does not hold then:
7: 𝑆 ∗ ← 𝑆 ∗ − {n𝑖 }
na

8: end if
9: end for
ur

10: if 𝑆 ∗ ≠ ∅ do
Jo

11: 𝜊 𝐿 ← 𝑇𝑟𝑢𝑒
12: n𝛼 ← argmax{𝑦̃𝑖 }
n𝑖 ∈𝑆 ∗

13: n𝛽 ← argmax{∆𝑡̃𝑖𝑗 }
n𝑗 ∈𝐴𝑖

14: n𝑃𝛼 ← n𝛽
15: Update {𝑡̃𝑘′ } for 𝑘 = {1,2, ⋯ , |𝐸|} with Equation (8, 9)
16: 𝑡𝑟𝑒𝑒(n𝛼 ) ← 𝑡𝑟𝑒𝑒(n𝛼 ) − {n𝛼 }
17: 𝑡𝑟𝑒𝑒(n𝛽 ) ← 𝑡𝑟𝑒𝑒(n𝛽 ) + {n𝛼 }
18: end if
2
3 While the utilization of Branch Grafting in redistribution will greedily select the
4 node that can reduce the maximum difference the most, the adoption of Leaf Grafting
5 will conservatively select the node that bears the least undertaking volume. Moreover,
6 Branch Grafting has a broader search range compared to the local search of Leaf

19
1 Grafting, giving it better global search capability. For these reasons, these two sub-
2 algorithms are employed in combination in reassignment (Algorithm 2.3).
3
Algorithm 2.3 DRF Reassignment
Input: 𝑇𝑖 for e𝑖 ∈E (Node-to-Tree Affiliation)
& n𝑃𝑖 for n𝑖 ∈ 𝑁 (Parent-Child Hierarchy)
𝑡̃𝑘 , 𝑅̃𝑘 for 𝑇𝑘 ∈ 𝐹 & 𝑦̃𝑖 for n𝑖 ∈ 𝑁 (Simulation Result)
Output: 𝑇𝑖 for e𝑖 ∈E (Node-to-Tree Affiliation after Reassignment)
& n𝑃𝑖 for n𝑖 ∈ 𝑁 (Parent-Child Hierarchy after Reassignment)
𝜏 𝐵 & 𝜏 𝐿 (Counter for Branch Grafting and Leaf Grafting)
1: Initialization 𝜏 𝐵 , 𝜏 𝐿 ← 0

f
̃ with Equation (3)
2: Calculate 𝑆𝑆

oo
2
̃ ≥ (𝜃 ∙ 𝑡̃̅) do
3: if 𝑆𝑆
4: Run Branch Grafting

r
5:
6:
if 𝜊 𝐵 do
𝜏𝐵 ← 1
-p
re
7: end if
̃ with Equation (3)
lP

8: Update 𝑆𝑆
2
̃ < (𝜃 ∙ 𝑡̃̅) do
9: while 𝑆𝑆
na

10: Run Leaf Grafting

11: if 𝜊 𝐿 do
ur

12: 𝜏𝐿 ← 𝜏𝐿 + 1
13: end if
Jo

14: Update 𝑆𝑆̃ with Equation (3)

15: end while
4
5 As shown above, one round of redistribution performs several Leaf Grafting after
6 performing at most one Branch Grafting. The specific number of both is influenced by
2
7 the maximum tolerable SSDM value (𝜃 ∙ 𝑡̃̅) where 𝜃 is a predefined upper limit of
8 relative range of {𝑡1̃ , 𝑡̃2 , ⋯ , 𝑡̃|𝐸| } . The DRF obtained from the reassignment is re-
9 evaluated by simulation, the result from which becomes the basis for the next round of
10 reassignment. As a termination criterion for this iterative optimization, the number of
11 executions of these sub-algorithms in each round is recorded with 𝜏 𝐵 and 𝜏 𝐿 ,
12 respectively. When 𝜏 𝐵 + 𝜏 𝐿 = 0, it means that there is no room for further optimization
13 via Grafting, and the final DRF is thus outputted.

14 3.4 Requirement of Evacuation Simulation

20
1 Cooperating with the above optimization method, a simulation model capable of
2 reproducing the crowd evacuation dynamics under the influence of signs is required.
3 Among the evacuation simulation models at different scales, the fine-grand microscopic
4 models offer significant advantages in depicting heterogeneous evacuees-sign
5 interactions, making them well-suited for simulation evaluation.

6 In addition to ensuring the reliability of the simulation model, accurately quantifying

7 the evacuee-sign interaction (i.e., reliable inputs of all 𝜀𝑖𝑗 ) can efficiently accelerate
8 optimization. These parameters are the rate at which evacuees in a sub-area change their
9 movement direction under the influence of signage compared to self-organized
10 evacuation. Regardless of the simulation model, the spontaneous escape direction of

f
the evacuees in a static environment without considering the person-to-person

oo
11
12 interaction can be represented by a vector field. While the directions of the vectors in

r
13 CA are determined by the steepest descent direction in the static field (Fig. 5 (a))
14
-p
whereas in SFM, it aligns with the direction of the resultant force of the desire force
and the obstacle repulsion (Fig. 5 (b)).
re
15
lP
na
ur
Jo

16
17 Fig. 5. Vector field in (a) CA and (b) SFM.

18 Building on research that quantifies the evacuees-sign interaction with perceiving

19 probability model (PPM) in the max covering area, the Monte Carlo method can be
20 employed for estimating 𝜀𝑖𝑗 . To prepare for the estimator, both spontaneous evacuation
21 directions according to the above discrete vector field and comply-with-sign directions
22 are generated at various locations within each subarea (shown in Fig. 6 (a) & (b)). When
23 the evacuees are uniformly distributed in the subarea, their movement patterns are
24 determined by a mixed vector field. In that field, the vector of each position is randomly
25 selected from two directions according to the probability derived from the PPM. By
26 generating the mixed vector field multiple times, the initial value of 𝜀𝑖𝑗 can be estimated

21
1 by averaging the flux ratio for flow in the direction of the boundary between n𝑖 and n𝑗
2 with the following formula.
T T
⃗ 𝑖𝑗 𝑉𝑖𝑡 𝐽𝑖 ⁄∑n𝑘 ∈ 𝐴𝑖 𝑢
∑𝜆𝑡=1 [𝑢 ⃗ 𝑖𝑘 𝑉𝑖𝑡 𝐽𝑖 ]
3 𝜀𝑖𝑗 = (13)
𝜆
4 where 𝜆 is the number of times the mixed vector field is generated. 𝑢
⃗ 𝑖𝑗 denotes the unit
5 normal vector of the b𝑖𝑗 whose positive direction is the direction of outflow from n𝑖 .
6 With the subareas obtained by the Voronoi partition, this vector can be derived using

7 the node positions of n𝑖 and n𝑗 . 𝑉𝑖𝑡 is a 𝑚 × 2 matrix, where each row corresponds to a

8 vector in the PPM mixed vector field obtained by the 𝑡-th probabilistic choice. Thus,
9 the number of rows 𝑚 is equal to the number of vectors in the subarea corresponding

f
oo
10 to n𝑖 determined by the level of discretization. 𝐽𝑖 is an 𝑚 × 1 all-one matrix.

r
-p
re
lP

11
na

12 Fig. 6. (a) Spontaneous, (b) Comply-with-sign and (c) PPM mixed vector field.

13 However, these calculated values may be prone to misestimated due to the

ur

14 assumption of perfect uniformity in the distribution of evacuees. As each round of

Jo

15 simulation results is generated, they can be updated with 𝜀𝑖𝑗 ← 𝑦̃𝑖𝑗 /(𝑥𝑖 + ∑𝑛𝑘∈𝐴𝑖 𝑦̃𝑘𝑖 ).

16 4. Numerical Test
17 To evaluate the effectiveness and efficiency of the method, we conducted numerical
18 tests in a university canteen during its operating hours. This scenario represents a typical
19 situation characterized by high crowd density and a complex internal layout. The venue,
20 located at Beihang University in Beijing, and its interior layout is depicted in Fig. 7. In
21 this figure, the light blue area represents the accessible area, which encompasses the
22 space outside of obstacles and the hypothetical fire-affected area indicated in red.
23 Additionally, we hypothetically installed 92 dynamic exit signs, denoted by green stars,
24 throughout the building. Given that the distribution of occupants can vary during the
25 observation period, we conducted tests considering four different crowd sizes, as
26 illustrated in Fig. 8.
27 While our proposed method can be combined with microsimulation models of
28 different mechanisms (as mentioned in Section 3.4), the model we adopted in numerical

22
1 tests is SFM. Unlike CA, which is not so prominent in current research hotspots, SFM's
2 algorithms have continuously improved according to current realistic research needs
3 and are frequently applied to reproduce guided evacuation dynamics[17,21,58]. Even
4 so, a common concern is that its computational cost balloons with crowd size, making
5 it impractical to work with metaheuristic-based optimization methods, particularly in
6 large buildings. Therefore, preliminary tests were conducted to determine the impact of
7 our proposed method on the feasibility of applying this type of model to optimize the
8 DSP. In the following tests, the effectiveness of the optimization method was reflected
9 in whether the maximum value or the relative range of {𝑡1̃ , 𝑡̃2 , ⋯ , 𝑡̃|𝐸| } obtained from
10 subsequent simulations decreases, while the times of simulation execution before an
11 acceptable near-optimal solution was regarded as the indicator of the method’s
12 optimization efficiency.

f
r oo
-p
re
lP
na
ur

13
Jo

14 Fig. 7. Internal layout of the canteen.

23
 f
r oo
-p
re
1
lP

2 Fig. 8. The distribution of occupants when the total number of them is

3 (a) 350, (b) 500, (c) 600, and (d) 750.
na

4 4.1 Impact of Crowd Size

ur

5 The initial numerical tests involved observing the optimization process for varying
Jo

6 crowd sizes. At this stage, it was assumed that all evacuees strictly adhere to the
7 instructions on the exit signs. By taking the different evacuee distributions depicted in
8 Fig. 8 as input and setting 𝜃 = 5% , we recorded the variations in the simulated
9 evacuation time {𝑡1̃ , 𝑡̃2 , 𝑡̃3 } and predicted evacuation time {𝑡1̃ ′ , 𝑡̃2′ , 𝑡̃3′ } for the three exits
10 𝑒1 , 𝑒2 , and 𝑒3 until the termination condition was met. As shown in Fig. 9, the
11 optimization process was labeled in the form of (𝑥, 𝑦, 𝑧), where 𝑥 denoted the number
12 of rounds of the current optimization iteration, while 𝑦 = 𝜏 𝐵 and 𝑧 = 𝜏 𝐿 were
13 respectively the number of Branch Grafting and Leaf Grafting already executed in the
14 current iteration. (Noted that when 𝑥, 𝑦 and 𝑧 were all 0, the {𝑡1̃ ′ , 𝑡̃2′ , 𝑡̃3′ } represented the
15 prediction results of the linear model in the DRF initialization algorithm, i.e., the value
16 in the temporary list 𝐿∗ in Algorithm 1, while when only 𝑥 ≠ 0 , the recorded
17 corresponded to the 𝑥-th simulation results.) On this basis, the average prediction error
18 (denoted by 𝛿) between {𝑡1̃ , 𝑡̃2 , 𝑡̃3 } and {𝑡1̃ ′ , 𝑡̃2′ , 𝑡̃3′ } at the end of each optimization round
19 (i.e., when 𝜊 𝐿 return 𝐹𝑎𝑙𝑠𝑒) can be calculated with the following equation.
24
 1 3 |𝑡̃𝑘 − 𝑡̃𝑘′ |
1 𝛿= ∑ (13)
3 𝑘=1 𝑡̃𝑘
2 The trend of these average prediction errors is shown with the red line in Fig. 9. And
3 Fig. 10 illustrates the DRF obtained through reassignment at the end of each
4 optimization round under the crowd size of 750.
5 It is noticeable that throughout the optimization process, the evacuation time of
6 each exit gradually approached equilibrium in successive simulations, resulting in a
7 consistent reduction in the total evacuation time (i.e., the maximum value among
8 {𝑡1̃ , 𝑡̃2 , 𝑡̃3 }). In all 4 scenarios, only 5-6 rounds of optimization were required before the
9 termination condition was reached, indicating that the computationally demanding
10 SFM only needed to be executed 5-6 times. In contrast to the relatively low-density

f
cases, although the Grafting algorithms were executed more frequently in each round

oo
11
12 of optimization in the high-density cases, the marginal increase in the number of

r
13 optimization rounds (from 5 rounds to 6 rounds) means that the computational cost of
14
-p
the entire process did not exhibit linear growth with the crowd size. The optimization
efficiency was primarily attributed to the significant decrease in the average prediction
re
15
16 error (from above 20% to below 5%) after obtaining the simulation output. The minor
lP

17 fluctuations prior to reaching the lowest value(1~2%) show that the reduction was not
18 achieved in one stride but by continuously learning the simulation feedback.
na
ur
Jo

25
 f
r oo
-p
re
lP
na
ur
Jo

1
2 Fig.9. Optimization process for crowd size of (a) 350, (b) 500, (c) 600, and (d) 750.

26
1

f
r oo
-p
re
lP
na
ur
Jo

2
3 Fig.10. The DRF obtained by reallocating at the end of the (a) 1st, (b) 2nd, (c) 3rd, (d) 4th, (e) 5th, and (f) 6th round under the population size of
4 750.

27
1 4.2 Impact of Crowd Distribution
2 To further investigate the subject, a second set of numerical tests was conducted to
3 compare the influence of the uniformity of the evacuee distribution on optimization
4 efficiency. Similarly, the four evacuee distributions shown in Fig. 8 were utilized but
5 averaged to varying extents using cells of different sizes (i.e., 5m×5m, 2m×2m, 1m×1m,
6 0.5m×0.5m). In particular, the evacuees’ locations were redistributed to the uniform
7 distribution within the cells. Consequently, the larger the averaged cells, the more
8 uniform the distribution of evacuees became. For each of the 16 combinations resulting
9 from pairing the four crowd sizes with the four grid sizes, diverse initial positions of
10 evacuees were generated based on 20 random grid phase differences. Furthermore, the

f
number of iteration rounds required for optimizing under all initial conditions was

oo
11
12 recorded and organized into groups, as shown in Fig. 11.

r
-p
re
lP
na
ur
Jo

13

14 Fig.11. Optimized efficiency under different evacuee distribution uniformity.

15 Upon analyzing the four data sets, it is apparent that the optimization efficiency
16 gradually decreases and becomes unstable at different crowd sizes. This is evidenced
17 by the rise in the average number of iteration rounds (from 0.25 to 0.8) and an increase
18 in the interquartile range (IR) reaching 1.15. This can be attributed to the fact that the
19 uniform distribution of evacuees aligns closely with the node-based statistical
20 assumption, which makes the linear prediction in the DRF initialization algorithm
21 closer to the simulation and facilitates the rapid learning of effective parameters in the
22 Grafting algorithm. Furthermore, we can find that at a crowd size of 350, despite the
28
1 maximum number of optimization rounds reaching 7, the IR stays at 0 across different
2 uniformities, with the mean value only increasing by 0.25. This suggests that this effect
3 is magnified as the crowd size increases. This is caused by the uneven distribution at
4 overall high densities, which tends to result in extremely crowded areas characterized
5 by intensified inter-individual interactions that emerge as complex dynamics, ultimately
6 leading to an escalation in prediction errors.

7 4.3 Impact of Network Scale

8 In addition to the total number of evacuees and their distribution uniformity, the
9 scale of the evacuation network, which is influenced by architectural factors, is another
10 aspect that impacts optimization efficiency. Consequently, the objective of the third set

f
of numerical experiments was to analyze the impact of the number of nodes in the

oo
11
12 network on optimization efficiency. During converting EN into DRF, exit nodes and

r
13 internal nodes play distinct roles. Therefore, we set more exits (i.e., 4, 5, 6) and signs
14
-p
(i.e., 128, 230) and initialized the initial distribution of evacuees according to the
scenario in Fig. 8 (d). For each combination of node sizes, the number of iterative
re
15
16 rounds needed for optimization is recorded in groups, as shown in Fig. 12.
lP
na
ur
Jo

17
18 Fig.12. Optimized efficiency at different node scales.

19 A significant trend is that the optimization efficiency decreases as the number of

20 internal nodes increases but increases as the number of exit nodes increases. For
21 instance, when there are 92 internal nodes and 3 exit nodes, the average number of
22 iteration rounds is 5.95. However, this value increases to 9.3 when the number of
23 internal nodes rises to 230, and decreases to 3.75 when the number of exit nodes doubles.
29
1 The ratio of these two factors signifies the average number of nodes on each tree in the
2 DRF. A greater number of root nodes results in the division of the EN into more DRTs,
3 frequently resulting in a decrease in average height. Conversely, an increase in internal
4 nodes has the opposite effect. A reduced height of DRT enables enhanced performance
5 of the initialized DRF, leading to more balanced evacuation times of exits in the initial
6 simulation and more frequent Leaf Grafting in each iteration. These factors contribute
7 to improving optimization efficiency when the prediction error is low. Moreover,
8 regarding the extent of change in optimization efficiency, there is an increase of 35% to
9 80% in the average number of iteration rounds when the number of internal nodes
10 increases to 2.5 times the previous number. Notably, the IQR remains constant at 1.75
11 in all three cases when the number of exit nodes is at its maximum value. This trend

f
oo
12 suggests that the method can effectively manage the rise in computational cost in multi-
13 exit facilities that necessitate a substantial number of dynamic signs for evacuation

r
14 guidance, making it a promising application in large buildings.
-p
15 4.4 Impact of Evacuee Compliance
re
16 In addition to the aforementioned factors related to problem size, the last set of tests
lP

17 explored the impact of evacuee's compliance, a psychological-level factor that cannot

18 be precisely determined in the system, on both the efficiency and validity of the
na

19 optimization. The scenarios in this set of tests employed the basic setup mentioned at
20 the beginning of this section. However, instead of assuming full obedience, a linearly
ur

21 decreasing obedience rate with increasing distance[4] was adopted. The number of
Jo

22 iterative rounds needed for optimization at different predefined obedience rates is

23 shown in Fig. 13.
24 Since the compliance rate is predetermined in the simulation model and can be
25 learned by the optimization method (see Section 3.3), it does not significantly impact
26 the optimization efficiency. With an increase in the compliance rate from 50% to 100%,
27 the stochasticity of the entire evacuation process decreases, ensuring the stability of the
28 optimization efficiency and significantly reducing the volatility of the number of
29 iteration rounds. Consequently, the IQR is reduced from 1.75 to 0. In addition, it is
30 noteworthy that even at 50% compliance, the range of the distribution of the number of
31 iterations is only 4 and exhibits a significant left-skew. Thus, it can be assumed that the
32 computational cost can still be maintained at a low level with a high probability in this
33 scenario.

30
 f
oo
1
2 Fig.13. Optimization efficiency under different compliance.

r
3 -p
Even so, as we cannot guarantee the accuracy of the compliance level being input
4 in the simulation module, robustness tests were conducted under the biased compliance
re
5 rates. After setting the direction according to the results obtained from the optimization
lP

6 at a specific compliance rate (50%, 60%, 70%, 80%, 90%, 100%), we adjusted the
7 compliance rate and reran the simulation module. Upon conducting additional
na

8 simulations, the average value of the output relative range, which had been controlled
9 to within 5% by the parameter 𝜃 in optimization, was recorded (as presented in Table
ur

10 2). Furthermore, the relationship between the distribution of this index and compliance
deviation (i.e., compliance rate in optimization minus compliance rate in test) is shown
Jo

11
12 in Fig. 14.

13 Table 2. The average relative range in the validation simulations under different
14 compliance.
optimization 50% 60% 70% 80% 90% 100%
validation compliance compliance compliance compliance compliance compliance
50%
3.854 6.874 8.745 16.09 16.137 17.934
compliance
60%
6.781 3.654 7.183 8.483 12.671 14.965
compliance
70%
7.85 5.946 2.78 6.512 8.934 13.528
compliance
80%
13.355 6.897 6.672 3.1235 7.312 7.673
compliance
90%
17.439 12.472 6.347 6.573 2.163 6.123
compliance
100%
22.034 14.463 12.405 7.673 5.446 1.674
compliance
15

31
 f
oo
1

r
2 Fig.14. Optimization effectiveness under different compliance deviations.
-p
3 An in-depth analysis of the two charts reveals that as the compliance deviation
re
4 changes from 0 to -50% and 50%, the mean value of the relative range of simulation
lP

5 output gradually increases from 2.63% to 21.23% and 18.00%, respectively. Likewise,
6 the IQR also increases from 1.66% to 6.52% and 3.51%, respectively. This suggests
na

7 that both overestimation and underestimation of compliance during the optimization

8 process can result in erratic and significant underperformance of the final plan.
ur

9 Furthermore, the diagonal trend observed in Table 2 demonstrates that while this trend
is symmetrical on both sides of the axis where the deviation is equal to 0, the increase
Jo

10
11 in relative range is smaller when the compliance is relatively high in both optimization
12 and validation, particularly when the deviation is 0. For instance, at a compliance rate
13 of 50%, the mean value of the relative range is 3.854%, which is significantly higher
14 than 1.674% in the fully compliant condition.
15 The implications of this set of tests at the application level are threefold: (1) It is
16 necessary to obtain relatively accurate compliance to enhance the validity of
17 optimization results through controlled experiments or evacuation drills; (2) In
18 situations where determining the precise level of compliance is challenging, opting for
19 intermediate values within the estimated range helps to control the bias caused by the
20 optimization results; (3) If the level of compliance can be enhanced through advocacy
21 efforts, the impact of compliance deviation on the effectiveness of optimization can be
22 mitigated to some extent.
23 Based on proposing this method, we explored the influence of different factors on

32
1 solving the DSP problem. Through the numerical tests, we found that increases in crowd
2 size and inhomogeneity, interior space scale, and crowd non-compliance would result
3 in varied degrees of increased computational burden. Conversely, adding more exits not
4 only decreases the evacuation time but also reduces the planning time. More
5 importantly, the proposed direction-setting method, embedded with the SFM,
6 maintained high optimization efficiency despite adverse factors. Specifically, the
7 number of optimization iterations just slightly increases when the scale of the factor
8 increases exponentially. This control over computational requirements is attributed to
9 the fact that the reassignment algorithm which incorporates performance predictions of
10 the enhanced DRF in the simulator based on network flow data output by the previous
11 simulation. In addition, this study provides a preliminary discussion of the impact of

f
oo
12 misestimated obedience on the validity of the optimization results and accordingly
13 suggests supplementary management tools.

r
14 5. Conclusion -p
re
15 In engineering practice, Automated Dynamic Exit Signs (ADES) are considered a
16 practical solution for improving the efficiency of building evacuation. However, ADES
lP

17 is still far from engineering practice due to the high computational cost of crowd
18 simulation that requires multiple executions. Although this study does not fully address
na

19 the computationally demanding challenge, it represents progress by minimizing the

computational cost in solving the simulation-based direction-setting problem. By
ur

20
21 leveraging the DRF-based encoding and dedicated algorithms, the proposed method
Jo

22 can reduce the times of simulation execution by around two orders of magnitude (the
23 problem at our test scale normally requires 300-500 simulation evaluations when solved
24 with metaheuristics like Genetic Algorithm). Since the method does not demand high
25 granularity in the simulation model, it would have greater potential for application once
26 faster simulation models are proposed.
27 A major feature of this method is its compatibility with different simulation models.
28 Even if the accuracy of the current simulation model may be questioned, the reliability
29 of the optimization results could be equally improved by combining optimization
30 modules with more reliable simulators. Similarly, if the simulation process, which
31 accounts for the majority of the computational time, can be accelerated, the
32 computational time required for the entire planning process will be reduced by nearly
33 the same amount. Specifically, we may need a data-driven model that can determine the
34 simulation output given the initial scene information of the evacuation (as shown in
35 Scheme I in Fig. 15). It is encouraging to see that relevant studies are being carried out,
33
1 and the model proposed has yielded reasonably satisfactory outcomes. However, for the
2 purpose of general discussion, this study still used the traditional model. Further, the
3 entire planning process could potentially be replaced by a machine learning model that
4 can directly provide optimized plans by analyzing the initial scene features (as shown
5 in the Scheme II in Fig. 15). When the training data for this real-time planning model
6 are needed, this method can substantially reduce the time required to obtain the
7 optimized direction setting plan for various initial scenarios. Given its potential for
8 engineering application potential, this method represents a significant step toward
9 ADES.

f
r oo
-p
re
lP
na
ur

10
Jo

11 Fig.15. Schematic diagram of two schemes for implementing ADES

12 While the approach proposed in this study lays a solid ground for realizing ADES,
13 three major extensions can further enhance its potential for engineering applications.
14 Firstly, current optimization methods do not consider the potential spread of hazards
15 and changes in the building passage. In future studies aiming at developing real-time,
16 instantaneous ADES, accelerated simulation will become a significant focus. One
17 potential technical route is to build neural network models that can directly output the
18 data needed for optimization. Secondly, when dealing with evacuation tasks in multi-
19 story buildings, the current method requires dividing them into multiple single-story
20 tasks. While such treatment can ensure optimization efficiency by reducing the number
21 of nodes through distributed computing, coordinating evacuation between stories from
22 a comprehensive managerial perspective remains challenging. A current solution would
23 be to execute the optimization method from the lower floors to the upper floors and

34
1 subsequently develop diversion schemes and guidance measures at the end of the
2 staircases based on the simulation results under the intervention of the output
3 optimization scheme. Lastly, to augment the optimization objectives, safety risks
4 encountered during the evacuees' escape could be considered alongside evacuation
5 efficiency. For this purpose, it is necessary to construct a bi-objective program or to
6 take the upper limit of the evacuation time as a constraint.
7

8 Acknowledgement
9 To acknowledge the funding support by the Research Grant Council Grant No:
10 CityU11216920 and the National Natural Science Foundation of China No: 72171006

f
oo
11

r
-p
re
lP
na
ur
Jo

35
1 Reference
2 [1] W. Wu, X. Zheng, A Systematic Analysis of Subgroup Research in Pedestrian and Evacuati
3 on Dynamics, IEEE Transactions on Intelligent Transportation Systems. (2023). https://doi.
4 org/10.1109/TITS.2023.3318417.
5 [2] A. Balboa, J. González-Villa, A. Cuesta, O. Abreu, D. Alvear, Testing a real-time intelligen
6 t evacuation guiding system for complex buildings, Saf Sci. 132 (2020). https://doi.org/10.1
7 016/j.ssci.2020.104970.
8 [3] Y. Huang, D. Cao, Decision response of subway evacuation signs based on brain componen
9 t features, Neural Comput Appl. 34 (2022) 6705–6719. https://doi.org/10.1007/S00521-021
10 -06150-Z/TABLES/7.
11 [4] Z. Zhang, L. Jia, Y. Qin, Optimal number and location planning of evacuation signage in pu
12 blic space, Saf Sci. 91 (2017) 132–147. https://doi.org/10.1016/j.ssci.2016.07.021.

f
oo
13 [5] R.K. Dubey, W.P. Khoo, M.G. Morad, C. Hölscher, M. Kapadia, AUTOSIGN: A multi-crite
14 ria optimization approach to computer aided design of signage layouts in complex building
15 s, Comput Graph. 88 (2020) 13–23. https://doi.org/10.1016/J.CAG.2020.02.007.

r
16 [6] R. Fan, Z. Dai, S. Tian, T. Xia, C. Huang, Research on spatial information transmission effi
-p
17 ciency and capability of safety evacuation signs, Journal of Building Engineering. 71 (202
re
18 3) 106448. https://doi.org/10.1016/J.JOBE.2023.106448.
19 [7] M. Zhou, H. Dong, P.A. Ioannou, Y. Zhao, F.-Y. Wang, Guided crowd evacuation: approach
lP

20 es and challenges; Guided crowd evacuation: approaches and challenges, IEEE/CAA Journ
21 al of Automatica Sinica . (2019). https://doi.org/10.1109/JAS.2019.1911672.
22 [8] M. Li, C. Xu, Y. Xu, L. Ma, Y. Wei, Dynamic Sign Guidance Optimization for Crowd Evac
na

23 uation considering Flow Equilibrium, J Adv Transp. 2022 (2022) 1–19. https://doi.org/10.1
24 155/2022/2555350.
ur

25 [9] D. Darsena, G. Gelli, I. Iudice, F. Verde, Sensing Technologies for Crowd Management, Ad
26 aptation, and Information Dissemination in Public Transportation Systems: A Review, IEE
Jo

27 E Sens J. 23 (2023) 68–87. https://doi.org/10.1109/JSEN.2022.3223297.

28 [10] X. Zhou, H. Li, J. Wang, J. Zhao, Q. Xie, L. Li, J. Liu, J. Yu, CloudFAS: Cloud-based build
29 ing fire alarm system using Building Information Modelling, Journal of Building Engineeri
30 ng. 53 (2022) 104571. https://doi.org/10.1016/J.JOBE.2022.104571.
31 [11] H. Fang, S. Lo, J.T.Y. Lo, Building Fire Evacuation: An IoT-Aided Perspective in the 5G E
32 ra, Buildings 2021, Vol. 11, Page 643. 11 (2021) 643. https://doi.org/10.3390/BUILDINGS
33 11120643.
34 [12] K. Saini, S. Kalra, S.K. Sood, Disaster emergency response framework for smart buildings,
35 Future Generation Computer Systems. 131 (2022) 106–120. https://doi.org/10.1016/J.FUT
36 URE.2022.01.015.
37 [13] H. Sharbini, R. Sallehuddin, H. Haron, Crowd evacuation simulation model with soft comp
38 uting optimization techniques: a systematic literature review, Journal of Management Anal
39 ytics. 8 (2021) 443–485. https://doi.org/10.1080/23270012.2021.1881924.
40 [14] H. Xie, Investigation into the interaction of people with signage systems and its implement
41 ation within evacuation models, (2011) Doctoral dissertation, University of Greenwich.
42 [15] N. Kaur, H. Kaur, A Multi-agent Based Evacuation Planning for Disaster Management: A N
43 arrative Review, Archives of Computational Methods in Engineering. 29 (2022) 4085–411

36
 1 3. https://doi.org/10.1007/S11831-022-09729-4/TABLES/6.
2 [16] H. Abdelgawad, B. Abdulhai, Emergency evacuation planning as a network design problem:
3 A critical review, Transportation Letters. 1 (2009) 41–58. https://doi.org/10.3328/TL.2009.
4 01.01.41-58.
5 [17] H. Liu, H. Chen, R. Hong, H. Liu, W. You, Mapping knowledge structure and research tren
6 ds of emergency evacuation studies, Saf Sci. 121 (2020) 348–361. https://doi.org/10.1016/J.
7 SSCI.2019.09.020.
8 [18] E. Bakhshian, B. Martinez-Pastor, Evaluating human behaviour during a disaster evacuatio
9 n process: A literature review, Journal of Traffic and Transportation Engineering (English E
10 dition). 10 (2023) 485–507. https://doi.org/10.1016/J.JTTE.2023.04.002.
11 [19] P. Elsayed, H. Mostafa, M. Marzouk, BIM based framework for building evacuation using
12 Bluetooth Low Energy and crowd simulation, Journal of Building Engineering. 70 (2023) 1
13 06409. https://doi.org/10.1016/J.JOBE.2023.106409.

f
14 [20] H. Vermuyten, J. Beliën, L. De Boeck, G. Reniers, T. Wauters, A review of optimisation mo

oo
15 dels for pedestrian evacuation and design problems, Saf Sci. 87 (2016) 167–178. https://do
16 i.org/10.1016/J.SSCI.2016.04.001.

r
17 [21] H. Sharbini, R. Sallehuddin, H. Haron, Crowd evacuation simulation model with soft comp
18
19
-p
uting optimization techniques: a systematic literature review, Journal of Management Anal
ytics. 8 (2021) 443–485. https://doi.org/10.1080/23270012.2021.1881924.
re
20 [22] H. Ran, L. Sun, X. Gao, Influences of intelligent evacuation guidance system on crowd eva
21 cuation in building fire, Autom Constr. (2013). https://doi.org/10.1016/j.autcon.2013.10.02
lP

22 2.
23 [23] G.S. Zhi, S.M. Lo, Z. Fang, A graph-based algorithm for extracting units and loops from ar
na

24 chitectural floor plans for a building evacuation model, Computer-Aided Design. 35 (2003)
25 1–14. https://doi.org/10.1016/S0010-4485(01)00171-3.
[24] H.W. Hamacher, S.A. Tjandra, Mathematical Modelling of Evacuation Problems: A State o
ur

26
27 f Art, Berichte Des Fraunhofer ITWM, Nr. 24 (2001). www.itwm.fhg.de.
Jo

28 [25] P. Lin, S.M. Lo, H.C. Huang, K.K. Yuen, On the use of multi-stage time-varying quickest ti
29 me approach for optimization of evacuation planning, Fire Saf J. 43 (2008) 282–290. https:
30 //doi.org/10.1016/J.FIRESAF.2007.08.005.
31 [26] P.H. Chen, F. Feng, A fast flow control algorithm for real-time emergency evacuation in lar
32 ge indoor areas, Fire Saf J. 44 (2009) 732–740. https://doi.org/10.1016/J.FIRESAF.2009.02.
33 005.
34 [27] G.J. Lim, S. Zangeneh, M. Reza Baharnemati, T. Assavapokee, A capacitated network flow
35 optimization approach for short notice evacuation planning, Eur J Oper Res. 223 (2012) 23
36 4–245. https://doi.org/10.1016/J.EJOR.2012.06.004.
37 [28] A. Desmet, E. Gelenbe, Capacity Based Evacuation with Dynamic Exit Signs, 2014 IEEE I
38 nternational Conference on Pervasive Computing and Communication Workshops . (2014)
39 332–337. https://doi.org/10.1109/PerComW.2014.6815227.
40 [29] M. Cisek, M. Kapalka, Evacuation Route Assessment Model for Optimization of Evacuatio
41 n in Buildings with Active Dynamic Signage System, Transportation Research Procedia. 2
42 (2014) 541–549. https://doi.org/10.1016/J.TRPRO.2014.09.094.
43 [30] J. Cho, G. Lee, S. Lee, An automated direction setting algorithm for a smart exit sign, Auto
44 m Constr. (2015). https://doi.org/10.1016/j.autcon.2015.05.004.

37
 1 [31] H. Kim, G. Lee, J. Cho, Prototype development and test of a server-independent smart exit
2 sign system: An algorithm, a hardware configuration, and its communication reliability, Au
3 tom Constr. 90 (2018) 213–222. https://doi.org/10.1016/J.AUTCON.2018.02.034.
4 [32] M. Fu, R. Liu, M. Rinker, BIM-based automated determination of exit sign direction for int
5 elligent building sign systems, Autom Constr. (2020). https://doi.org/10.1016/j.autcon.2020.
6 103353.
7 [33] F. Mirahadi, B.Y. McCabe, EvacuSafe: A real-time model for building evacuation based on
8 Dijkstra’s algorithm, Journal of Building Engineering. 34 (2021) 101687. https://doi.org/1
9 0.1016/J.JOBE.2020.101687.
10 [34] X. Pan, C.S. Han, K. Dauber, K.H. Law, Human and social behavior in computational mod
11 eling and analysis of egress, Autom Constr. 15 (2006) 448–461. https://doi.org/10.1016/J.A
12 UTCON.2005.06.006.
13 [35] H. Abdelgawad, B. Abdulhai, Emergency evacuation planning as a network design problem:

f
14 A critical review, Transportation Letters. 1 (2009) 41–58. https://doi.org/10.3328/TL.2009.

oo
15 01.01.41-58.
16 [36] Z. Gao, Y. Qu, X. Li, J. Long, H.J. Huang, Simulating the dynamic escape process in large

r
17 public places, Oper Res. 62 (2014) 1344–1357. https://doi.org/10.1287/opre.2014.1312.
18
19
[37]
-p
X. Wang, C. Mohcine, J. Chen, R. Li, J. Ma, Modeling boundedly rational route choice in c
rowd evacuation processes, Saf Sci. 147 (2022) 105590. https://doi.org/10.1016/J.SSCI.202
re
20 1.105590.
21 [38] A. Abdelghany, K. Abdelghany, H. Mahmassani, W. Alhalabi, Modeling framework for opt
lP

22 imal evacuation of large-scale crowded pedestrian facilities, Eur J Oper Res. 237 (2014) 11
23 05–1118. https://doi.org/10.1016/J.EJOR.2014.02.054.
na

24 [39] I. Park, G.U. Jang, S. Park, J. Lee, Time-dependent optimal routing in micro-scale emergen
25 cy situation, in: Proceedings - IEEE International Conference on Mobile Data Managemen
t, 2009: pp. 714–719. https://doi.org/10.1109/MDM.2009.122.
ur

26
27 [40] A. Abdelghany, K. Abdelghany, H. Mahmassani, A hybrid simulation-assignment modeling
Jo

28 framework for crowd dynamics in large-scale pedestrian facilities, Transp Res Part A Polic
29 y Pract. 86 (2016) 159–176. https://doi.org/10.1016/J.TRA.2016.02.011.
30 [41] J.C. Chu, A.Y. Chen, Y.F. Lin, Variable guidance for pedestrian evacuation considering con
31 gestion, hazard, and compliance behavior, Transp Res Part C Emerg Technol. 85 (2017) 66
32 4–683. https://doi.org/10.1016/J.TRC.2017.10.009.
33 [42] K. Wang, W. Yuan, W. Liang, Y. Yao, An optimal guidance strategy for fire evacuations: A
34 hybrid modeling approach, Journal of Building Engineering. 73 (2023) 106796. https://doi.
35 org/10.1016/J.JOBE.2023.106796.
36 [43] X. Chen, M. Treiber, V. Kanagaraj, H. Li, Social force models for pedestrian traffic–state o
37 f the art, Transp Rev. 38 (2018) 625–653. https://doi.org/10.1080/01441647.2017.1396265.
38 [44] Y. Li, M. Chen, Z. Dou, X. Zheng, Y. Cheng, A. Mebarki, A review of cellular automata mo
39 dels for crowd evacuation, Physica A: Statistical Mechanics and Its Applications. 526 (201
40 9) 120752. https://doi.org/10.1016/J.PHYSA.2019.03.117.
41 [45] P.B. Luh, C.T. Wilkie, S.-C. Chang, K.L. Marsh, N. Olderman, S.-C. Chang, IEEE TRANS
42 ACTIONS ON AUTOMATION SCIENCE AND ENGINEERING 1 Modeling and Optimi
43 zation of Building Emergency Evacuation Considering Blocking Effects on Crowd Movem
44 ent, IEEE Transactions on Automation Science and Engineering. 9 (2012). https://doi.org/1

38
 1 0.1109/TASE.2012.2200039.
2 [46] E.M. Cepolina, A methodology for defining building evacuation routes, Civil Engineering a
3 nd Environmental Systems. 22 (2005) 29–47. https://doi.org/10.1080/10286600500049946.
4 [47] X. Zong, S. Xiong, Z. Fang, W. Lin, Multi-objective ant colony optimization model for em
5 ergency evacuation, in: Proceedings - 2010 6th International Conference on Natural Compu
6 tation, ICNC 2010, 2010: pp. 2774–2778. https://doi.org/10.1109/ICNC.2010.5583265.
7 [48] C. Xie, D.Y. Lin, S. Travis Waller, A dynamic evacuation network optimization problem wi
8 th lane reversal and crossing elimination strategies, Transp Res E Logist Transp Rev. 46 (2
9 010) 295–316. https://doi.org/10.1016/J.TRE.2009.11.004.
10 [49] M.N.A. Khalid, U.K. Yusof, Dynamic crowd evacuation approach for the emergency route
11 planning problem: Application to case studies, Saf Sci. 102 (2018) 263–274. https://doi.org
12 /10.1016/j.ssci.2017.10.024.
13 [50] J. Kang, I.J. Jeong, J.B. Kwun, Optimal facility–final exit assignment algorithm for buildin

f
14 g complex evacuation, Comput Ind Eng. 85 (2015) 169–176. https://doi.org/10.1016/J.CIE.

oo
15 2015.03.012.
16 [51] Y. Li, W. Cai, A.A. Kana, Design of level of service on facilities for crowd evacuation usin

r
17 g genetic algorithm optimization, Saf Sci. 120 (2019) 237–247. https://doi.org/10.1016/J.S
18
19 [52]
SCI.2019.06.044.
-p
Y. Li, Y. Zhang, X. Pan, J.E. Taylor, BIM-based determination of indoor navigation sign la
re
20 yout using hybrid simulation and optimization, Autom Constr. 139 (2022) 104243. https://d
21 oi.org/10.1016/J.AUTCON.2022.104243.
lP

22 [53] L. Filippidis, E.R. Galea, S. Gwynne, P.J. Lawrence, Representing the influence of signage
23 on evacuation behavior within an evacuation model, Journal of Fire Protection Engineerin
na

24 g. 16 (2006) 37–73. https://doi.org/10.1177/1042391506054298.

25 [54] M. Fu, R. Liu, An approach of checking an exit sign system based on navigation graph net
works, Advanced Engineering Informatics. 46 (2020) 101168. https://doi.org/10.1016/J.AE
ur

26
27 I.2020.101168.
Jo

28 [55] R.S. Barr, R. V Helgaon, J.L. Kennington, Minimal spanning trees: An empirical investigat
29 ion of parallel algorithms, 1989.
30 [56] L.D. Vanumu, K. Ramachandra Rao, G. Tiwari, Fundamental diagrams of pedestrian flow c
31 haracteristics: A review, European Transport Research Review. 9 (2017). https://doi.org/10.
32 1007/s12544-017-0264-6.
33 [57] R.L. Francis, A “Uniformity Principle” for Evacuation Route Allocation*, n.d.
34 [58] X. Yang, X. Yang, Q. Wang, Y. Kang, F. Pan, Guide optimization in pedestrian emergency e
35 vacuation, Appl Math Comput. 365 (2020) 124711. https://doi.org/10.1016/J.AMC.2019.12
36 4711.
37

38

39
1 Appendix
2 Table A1
Symbol Representation
n𝑖 The i-th node in the evacuation network
𝑁 The set of nodes in the evacuation network
e𝑘 The k-th exit node in the evacuation network
𝐸 The set of exit nodes in the evacuation network
s𝑖 The subarea corresponding to n𝑖
𝑏𝑖𝑗 The boundary between s𝑖 and s𝑗
𝐴𝑖 The set of accessible neighbor nodes of n𝑖

f
𝐷 The weighted adjacency matrix of the evacuation network

oo
𝑥𝑖 The number of evacuees initially located in s𝑖
𝑑𝑖.𝑘 The geodesic distance from n𝑖 to e𝑘

r
𝑟𝑘 The evacuation capacity of e𝑘
𝑣
𝑇𝑘
-p
The average velocity of evacuees
The set of the nodes in the k-th tree
re
𝐿𝑘 The set of the leaf nodes in the k-th tree
lP

𝐵𝑘 The set of the branch nodes in the k-th tree

n𝑃𝑖 The parent node of n𝑖
The set including all child nodes of n𝑖
na

𝐶𝑖
𝑡̃𝑘 The total evacuation time of the k-th exit/tree in the simulation
𝑡̃ F The total evacuation time of the building/forest in the simulation
ur

𝑦̃𝑖𝑗 The total number of evacuees flowing from n𝑖 to n𝑗 in the simulation

Jo

𝑦̃𝑖 The total number of evacuees flowing out from n𝑖 in the simulation
𝑅̃𝑘 The time series of outflows from e𝑘 by time periods in the simulation
̃
𝑆𝑆 The sum of squares of deviations from mean of {𝑡1̃ , 𝑡̃2 , ⋯ 𝑡̃|𝐸| }
∆𝑡𝑖𝑗̃ The difference between the evacuation time of the tree to which n𝑖
belongs and that of n𝑗
∆𝑡̃𝑖 The maximum value of ∆𝑡̃𝑖𝑗 for n𝑗 ∈ 𝐴𝑖
′
𝑡̃𝑘 The estimated new value of 𝑡̃𝑘 when certain node is assumed to be
reassigned
𝜀𝑖𝑗 The estimated proportion of 𝑦̃𝑖 that flow into n𝑗 when n𝑃𝑖 = n𝑗
𝑦̃𝑖𝑚𝑎𝑥 The maximum value of undertaking volume for n𝑖 ’s reassignment
𝑜𝑝𝑡
𝑦̃𝑖 The optimal value of undertaking volume for n𝑖 ’s reassignment
𝑜𝑝𝑡
𝜂̃𝑖 The proximity of 𝑦̃𝑖 to 𝑦̃𝑖
𝜊𝐵 The observer for Branch Grafting execution
𝜏𝐵 The number of times executions Branch Grafting is executed
𝐿
𝜊 The observer for Leaf Grafting execution

40
 𝜏𝐿 The number of times executions Leaf Grafting is executed
𝜃 The maximum tolerable relative range of {𝑡1̃ , 𝑡̃2 , ⋯ 𝑡̃|𝐸| } (pre-set for
optimization)
𝑢
⃗ 𝑖𝑗 The unit normal vector of the b𝑖𝑗
𝜆 The number of times the mixed vector field is generated (pre-set for 𝜀𝑖𝑗
estimation)
𝑉𝑖𝑡 The matrix containing all vectors in the PPM mixed vector field
generated for the 𝑡-th time in s𝑖
𝐽𝑖 The all-one matrix for 𝜀𝑖𝑗 initialization
𝑖𝑛𝑑() The function for getting the subscript of the node in 𝑁
𝑑𝑖𝑠() The function for calculating the geodesic distance between two nodes
𝑡𝑟𝑒𝑒() The function for getting the tree that the node belongs to

f
oo
𝑡𝑘 () The function for predicting the evacuation time of 𝑇𝑘
𝐹() The function for extracting the average outflow rate under congestion

r
conditions from 𝑅̃

1
-p
re
lP
na
ur
Jo

41
Highlights
• Solved the Direction Setting Problem with the Directed Rooted Forest decomposition
• Designed dedicated optimization algorithms which leverage evacuation simulation output
• Considered the influence of compliance level in the direction setting planning
• Greatly improved the planning efficiency of the simulation-based direction setting
• Evaluated the method’s control to the combinatorial explosion of evacuation system

f
r oo
-p
re
lP
na
ur
Jo
Declaration of interests

☒ The authors declare that they have no known competing financial interests or personal relationships
that could have appeared to influence the work reported in this paper.

☐ The authors declare the following financial interests/personal relationships which may be considered
as potential competing interests:

of
ro
-p
re
lP
na
ur
Jo