230 likes | 560 Views
Improved Crowd Control Utilizing a Distributed Genetic Algorithm. John Chaloupek December 3 rd , 2003. Overview. Why Crowd Control? “Distributed” Genetic Algorithm? Goals Distributed Design GA Design & Representation Results Future Work. Crowds. Bad Stuff happens Fires
E N D
Improved Crowd Control Utilizing a Distributed Genetic Algorithm John Chaloupek December 3rd, 2003
Overview • Why Crowd Control? • “Distributed” Genetic Algorithm? • Goals • Distributed Design • GA Design & Representation • Results • Future Work
Crowds • Bad Stuff happens • Fires • Terrorist Attacks • Weapons of Mass Destruction • Natural Disasters
Crowds • People act irrationally in a disaster. • Panic • Confusion • Crowds often make the situation worse. • Sometimes the crowd is more dangerous than the disaster.
Crowd Control • First Responders (Police, Fire Dept., etc.) have limited capabilities to deal with crowds. • Barriers • Riot gear
Why use an EA? • Doable • Few other ways exist to simulate crowd behavior. • Can test new methods and ideas before putting them to work in a genuine situation.
Why use an EA? • Novel Methods • EA’s can help gather support for new methods that have yet to be proven effective. • Unexpected Discoveries • Could come up with methods that haven’t been thought of before.
“Distributed” GA? • Actually more of a Client/Server model. • Fitness evaluation is the most computationally intensive part of real world sized problems. • Fitness evaluations can be done in parallel, on multiple processors or multiple machines.
Similar Distributed Projects • Distributed.Net • Cryptography, Optimal Golomb Rulers • Seti@home • Signal Analysis • United Devices • Protein Modeling
Goals • See if a system for simulating crowd behavior & crowd control using a GA can be developed. • Reduce (virtual) fatalities. • Do it all in a reasonable amount of time.
Client/Server Model • Server runs GA and passes out members of the population to be evaluated. • Clients evaluate fitness.
GA Design Highlights • Rank based selection • Rank based competition (w/Elitist) • Uniform crossover • User specifiable parameters • Pc, Pm, Steepness of • Pop Size, #of Gens to run, How often to log,
GA Design Highlights • User specifiable parameters • Pc, Pm • Steepness of the Rank based probabilities. • Can set independently for selection and competition. • Pop Size, #of Gens to run • How often to log • Can specify a RNG Seed
Representation - Map • Walls, Exits and Damage sources (fires, chemical spills, etc.) are loaded from a BMP file.
Representation - Members • Members consist of what actions could be taken to control a crowd. • Place barricades • Set up noise sources • Direct people away from the scene
Evaluation • Simplistic AI “victims” are randomly placed on the scene. • Panic • Shortest Route to exit • Run away from most damage/noise • Follow the crowd • Try to pick proportions to most accurately simulate real situation.
Fitness • As victims remain on the scene, and fail to get away from sources of damage, they become hurt. • Fitness is the average of the health of the victims.
Results • 23.6% Improvement in 100 generations. • Pop Size: 1000 • B of Selection: 2 • B of Competition: 2 • Prob. Crossover: .2 • Prob. Mutation: .2
Summary • Client/Server code not working all that great. • Lots of room to expand in the future. • Surprisingly good results for what’s currently running.