Elsevier

Applied Soft Computing

Volume 66, May 2018, Pages 90-103
Applied Soft Computing

Guide them through: An automatic crowd control framework using multi-objective genetic programming

https://doi.org/10.1016/j.asoc.2018.01.037Get rights and content

Highlights

  • We propose an automatic framework based on genetic programming to generate crowd control strategies.

  • We propose a Multi-Objective Cartesian Genetic Programming (MO-CGP) algorithm capable of evolving not only parameters but also rule structures so as to find optimal control strategies for multi-objective optimization.

  • We demonstrate the effectiveness of the framework in a real life crowd control scenario to guide approximately 400 participants to pass through a multi-story building with shorter travel time and reduced congestion at hotspot segments, comparing to several common baseline crowd models.

  • We propose a rule analysis method through feature selection machine to identify a limited number of most important features in the evolved rules from GP algorithm effectively.

Abstract

We propose an automatic crowd control framework based on multi-objective optimisation of strategy space using genetic programming. In particular, based on the sensed local crowd densities at different segments, our framework is capable of generating control strategies that guide the individuals on when and where to slow down for optimal overall crowd flow in realtime, quantitatively measured by multiple objectives such as shorter travel time and less congestion along the path. The resulting Pareto-front allows selection of resilient and efficient crowd control strategies in different situations. We first chose a benchmark scenario as used in [1] to test the proposed method. Results show that our method is capable of finding control strategies that are not only quantitatively measured better, but also well aligned with domain experts’ recommendations on effective crowd control such as “slower is faster” and “asymmetric control”. We further applied the proposed framework in actual event planning with approximately 400 participants navigating through a multi-story building. In comparison with the baseline crowd models that do no employ control strategies or just use some hard-coded rules, the proposed framework achieves a shorter travel time and a significantly lower (20%) congestion along critical segments of the path.

Introduction

Crowd modelling and simulation has gained increasing attention from industry, academia and government due to its wide applications [2] to understand, replicate and predict crowd dynamics in various situations. As a natural extension to and an application of crowd modelling and simulation, crowd control aims to intervene [3] the movement of crowds in a desired manner so that certain objectives are met, for instance, to prevent turbulence or stampede in events involving massive crowds, to avoid bottlenecks of crowd flow, or to minimize overall travelling time, etc.

To apply appropriate crowd control strategies to intervene the crowd in a desired manner, one needs to first understand the implicit (unintervened) crowd dynamics under specific scenarios, which can be studied through crowd modeling and simulation. One promising approach is agent-based modelling (ABM), which treats individuals as agents that can perceive, decide and act independently based on some rules [4]. From the ABM perspective, crowd dynamics emerge from the motions of individuals, and the motions can be generated through a simplified two-layer movement model [5]. At the path planning layer (the higher layer), an agent plans/finds a path to navigate through the environment. The path segments are usually formed as a list of waypoints representing important landmarks and accessable areas. While at the collision avoidance layer (the lower layer), it avoids collisions with others while moving along the planned path. From the modeling perspective, there are some established methods to specify the rules for agents at both layers. For path planning, both shortest path algorithms (such as A*) [6] and accumulative segment-based algorithms that take account of vision range [7] have been well established to guide an agent to move through a set of static obstacles in an environment. For collision-avoidance, algorithms such as reciprocal velocity obstacle (RVO) and its variants [8], and social force model (SFM) [9] are proven efficient and widely adopted. With the two layers of movement behaviours, ABM can generate various crowd dynamics given the initial configurations of the agents (e.g., preferred speed, personal space factor etc.) and the environment (e.g., waypoints of paths, obstacles etc.).

Due to the complex interactions among the agents, the stochastic nature of the crowd model and the large number of parameters involved, finding a “good” crowd control strategy that explicitly intervenes movements of crowds in order to produce the “desired” crowd movements often requires a large number of simulations, which is time-consuming if performed manually. It is therefore important to automate the search process for optimal crowd control strategies. Evolutionary algorithms (EAs) are population-based non-deterministic search algorithms, which can be used to adaptively evolve a simulation model through automating the calibration of model parameters as well as model structures (such as behavioural rules of agents) [10], [11]. Although there are scattered existing works on using EAs for automatic crowd control, they mainly focus on the optimization of parameters, which may limit the search space by the fixed number of parameters. In this paper, we apply Genetic programming (GP) to enable both parameter and structure evolving for automatic crowd control, which will be depicted in Section 4.

The need for multi-objective optimization [12] is also essential for crowd control, as a good control strategy often needs to achieve different aspects of crowd dynamics simultaneously. For example, increasing the speed of an escalator may improve the flow rate of one segment of a path, while it may cause congestion at other (e.g., the subsequent) segments if there are spatial bottlenecks. Thus, the overall flow rate and the congestion conditions along the path need to be considered simultaneously in searching for a good crowd control strategy in this case. In this paper, our proposed GP-based framework can automatically search for the optimized parameters and rules used in an agent-based crowd model for crowd control purposes, specifically to optimize multiple objectives from the crowd dynamics perspective.

The rest of the paper is organized as follows: Section 2 describes the existing efforts in applying EAs to calibrate crowd simulation models, and traditional crowd control approaches. The problem of automatic crowd control through optimization of an agent-based model is formally defined in Section 3. As the proposed solution to address the problem, the GP-based crowd control framework is discussed in Section 4. In Section 5, we test the framework with two scenarios, a well studied evacuation scenario in [1] and a real life event planning scenario, where approximately 400 delegates are directed to leave a multi-story building with escalators transporting between stories. Section 6 concludes the paper and gives recommendations for future work.

Section snippets

Application of evolutionary algorithms in crowd simulation models

Modeling and simulation has become a promising approach to study crowd dynamics in recent years. Various models [4], [5], [13], [14] have been proposed with different focuses on particular aspects of a crowd according to the requirements of an application. One common and critical objective of these models is to generate realistic crowd behaviors through model calibration and validation [10], [11], [15], and variations of EAs have been applied to achieve this goal.

The most common idea is to use

Problem definition

In this paper, crowd control is defined as a multi-objective optimization problem over an agent-based crowd simulation model. The inputs of the crowd model consist of agent-specific configurations and scenario-specific configurations as shown in Fig. 1.

Agent-specific configurations include parameters describing an agent's characteristics (e.g., preferred walking speed and personal space factor etc.), as well as behavioural rules that determine its movement implicitly. In the two-layer motion

Proposed method

To address the crowd control optimization problem, we propose an automatic crowd control framework as shown in Fig. 1, with mainly two components: the simulation engine and the optimization system. The simulation engine and the optimization system form a feedback system: the control strategies generated by the optimization system are fed into the simulation engine to affect the crowd movement; while the output of the simulation engine (for instance, travel time and crowd density) are used by

Experiment studies

In theory, the proposed crowd control framework can be applied in any crowd control scenario, where scenario-specific parameters and the condition of control rules can be represented as a combination of measurable status of different components. In this paper, such components refer to individual segments’ densities. To evaluate the effectiveness of the proposed framework, we test it in two scenarios: first, we use a fundamental, and well-studied evacuation scenario as reported in [1] to find

Conclusions and future work

In this paper, we have proposed an automatic crowd control framework based on genetic programming. The framework consists of a simulation engine to generate crowd dynamics and an optimization system based on genetic programming. The framework aims to provide the “optimal” crowd control strategies in order to generate crowd dynamics with multiple objectives through simultaneous evolution of scenario-specific parameters and control rules that affect the agent's movement explicitly in the

Acknowledgements

Nan Hu and Wentong Cai would like to acknowledge the support from IHPC-NTU Joint R&D Project on “Symbiotic Simulation and Video Analysis of Crowds”; Jinghui Zhong is supported by the National Natural Science Foundation of China (Grant no. 61602181) and Fundamental Research Funds for the Central Universities (Grant no. 2017ZD053). The authors would like to thank Dr. Tan Singkuang and Dr. Yi Wenchao for their support and contribution to the paper.

References (35)

  • J. Zhong et al.

    Density-based evolutionary framework for crowd model calibration

    J. Comput. Sci.

    (2015)
  • D. Helbing et al.

    Simulating dynamical features of escape panic

    Nature

    (2000)
  • S.J. Guy et al.

    Modeling collision avoidance behavior for virtual humans

    Proceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems, vol. 2

    (2010)
  • D. Grieger

    An overview of crowd control theory and considerations for the employment of non-lethal weapons, Tech. Rep. DSTO-GD-0373, Land Operations Division System Sciences Laboratory, Australian Government, DSTO, Australia

    (2004)
  • W. Shao et al.

    Autonomous pedestrians

    Proceedings of the 2005 ACM SIGGRAPH/Eurographics symposium on Computer Animation

    (2005)
  • N. Hu et al.

    A pattern-based modeling framework for simulating human-like pedestrian steering behaviors

    Proceedings of the 19th ACM Symposium on Virtual Reality Software and Technology

    (2013)
  • M. Ghallab et al.

    Aepsilon – an efficient near admissible heuristic search algorithm

    Proceedings of the 8th International Joint Conference on Artificial Intelligence. Karlsruhe, FRG

    (1983, August)
  • S.J. Rymill et al.

    Psychologically-based vision and attention for the simulation of human behaviour

    Proceedings of the 3rd International Conference on Computer Graphics and Interactive Techniques in Australasia and South East Asia

    (2005)
  • J. Van den Berg et al.

    Reciprocal velocity obstacles for real-time multi-agent navigation

    Proceedings of the 2008 IEEE International Conference on Robotics and Automation

    (2008)
  • D. Helbing et al.

    Social force model for pedestrian dynamics

    Phys. Rev. E

    (1995)
  • J. Zhong et al.

    Automatic rule identification for agent-based crowd models through gene expression programming

    Proceedings of the 2014 International Conference on Autonomous Agents and Multi-Agent Systems

    (2014)
  • K. Deb

    Multi-Objective Optimization Using Evolutionary Algorithms

    (2001)
  • M. Moussaid et al.

    How simple rules determine pedestrian behavior and crowd disasters

    Proc. Natl. Acad. Sci. U. S. A.

    (2011)
  • S.J. Guy et al.

    Least-effort trajectories lead to emergent crowd behaviors

    Phys. Rev. E: Stat. Nonlinear Soft Matter Phys.

    (2012)
  • A. Lerner et al.

    Crowds by example

    Computer Graphics Forum, vol. 26

    (2007)
  • A. Johansson et al.

    Specification of the social force pedestrian model by evolutionary adjustment to video tracking data

    Adv. Complex Syst.

    (2007)
  • T.-Y. Li et al.

    Autonomous robots and agents

    An Evolutionary Approach to Crowd Simulation

    (2007)
  • Cited by (10)

    • A review on crowd simulation and modeling

      2020, Graphical Models
      Citation Excerpt :

      Firstly, when an emergency occurs, it is difficult to obtain the actual escape behavior of the crowd, especially in large-scale emergency evacuations. To re-construct such an emergency situation, researchers try to collect trajectory data through real-world experiments [160]. However, the experimental scene (such as the corridor and the hall) is too simple and small, which could not reflect the real-world situations.

    • Recent Advances in Crowd Movement Space Design and Control using Crowd Simulation

      2022, Transactions of the Japanese Society for Artificial Intelligence
    View all citing articles on Scopus
    View full text