A non-deterministic multiagent approach for emergency evacuation simulations

Presentation for IJUP 2017

Transcript

1. A non-deterministic multi-agent approach for emergency evacuation simulations Agents and Distributed Artificial Intelligence - 2016/2017 Gil Domingues / Pedro Pontes
2. Outline • Introduction • Problem Context • Proposal • Implementation • Experimentation • Conclusions • Future Work
3. Introduction
4. Introduction “An agent is anything that can be viewed as perceiving its environment through sensors and acting upon that environment through effectors.” Artificial intelligence : a modern approach by Stuart Russell, Peter Norvig “ Multiagent systems are made up of multiple interacting intelligent agents (…) able to cooperate, compete, communicate, act flexibly and exercise control over their behavior within the frame of their objectives” Multiagent Systems. Retrieved from https://mitpress.mit.edu/books/multiagent-systems-0
5. Introduction • Evacuating a space is a challenge, no matter the circumstances. • Analyzing succesful and unsuccesful evacuation attempts is necessary do develop an adequate evacuation plan.
6. Context • Known approaches can be based on fluid dynamics, on automatons or on agents. • Both the fluid dynamics and the automatons based approaches fail to consider the non deterministic nature of human behaviour.
7. Proposal • Create a multiagent system capable of simulating evacuation scenarios. • Model agent’s behaviour according to various agent attributes. • Capable of running different evacuation scenarios in different environments.
8. Implementation Details
9. Technologies used • Repast Simphony • JADE • SAJaS
10. Agent Attributes • Age • Area Knowledge • Independence • Altruism • Mobility • Panic • Patience
11. Agent Behaviours
12. Scream • An agent with an high level of panic is prone to screaming, alerting everyone in its neighbourhood and increasing their panic level. • Affecting attributes: panic.
13. Push • Agents with an high level of panic are prone to pushing any agent that is blocking its desired path. A push increases the level of panic and decreases the level of mobility of the pushed agent. • Affecting attributes: panic.
14. Help • In certain circumstances, an agent may help a injured agent by carrying it to the exit. • The injured agent can choose the “best” help if there are more than one agent offering help. • Affecting attributes: mobility, altruism and panic.
15. Sharing Directions • An agent may request directions from neighbouring agents in order to increase its area knowledge. • Affecting attributes: area knowledge, altruism.
16. Movement • It’s the agent’s goal to move towards the exit. • Can be divided into four possibilities, in order of priority: best path, visible exit, following crowds and random movement. • Affecting attributes: mobility, panic, area knowledge, patience and independence.
17. Customization • Scenario definition file (agent number and characteristics) • Environment definition file (evacuation site plant). Example of a small evacuation site
18. Experimentation and Results
19. Experimentation • Goal: analyse behaviour in specific conditions and confront it with the expected results.
20. Experimentation example (area knowledge) High area knowledge Low area knowledge
21. Experimentation example (choke points) Ample exits Limited exits
22. Experimentation example (overcrowding) Large population in a confined space
23. Conclusions
24. Conclusions • A working proof of concept was developed that allows the analysis of the interaction between agents in a configurable evacuation scenario. • Phenomena like clogging and herding (crowd dynamics) was observed. • Importance of area knowledge was observed.
25. Future work • Obstacles (fire, smoke, etc.) • Recognition of previously explored paths • Visual representation of a behavioural network