julian-agents

EXIT = Way Out
Julian Dymacek
April 29
Escape Panic Paper
• Dr. Dirk Helbing, Illes J. Farkas, Dr. Tamas Vicsek
• Point mass simulation
• Uses psychological forces to keep agents apart and
away from walls
• Uses friction to simulate the clogs in front of
doors
• Found a combination of rushing to doors and
following neighbors demonstrated escape panic
Craig Reynolds
• Craig Reynolds – Boids
– Separation, Alignment, Cohesion
• Craig Reynolds – Steering Behaviors
– Obstacle avoidance
– Wandering
– Following
Wander Behavior
What do I want to do?
• Reproduce the escape panic simulation
• Allow agents to be controlled by behaviors
not included in the escape panic paper
• Find behaviors that help agents quickly exit
from a room
Behaviors
• Closest
– Distance to door/ max distance
• Follow your neighbors
– Density of surrounding agents (agent area/ circle area)
• Go with the flow
– Avg speed of agents in radius no return/ max speed
• Popularity
– Density of agents around door (agent area/ half circle area)
Chromosome for GA
• Each behavior is a 5 bit gene
• Wander is the default behavior
• Another 5 bit gene represents the order of
applying behaviors
• A final 5 bit gene encodes desired speed
• 30 total bits
Tests
• Solved for best strategy with a single agent
and multiple agents
• Varied the percentage of agents who follow
neighbors with 95%, 75%, 50%, 25% and
0%
• Used two separate distributions of agents
• Agents had 20 seconds to escape
• Fitness was 1-(time to escape/ 20)
Results
• The Good
– Found ways besides go to closest
• The Bad
– Mostly found go to closest
• The Ugly
– The multi-agent tests could become inflated
The Good
Speed
Closest
Density
Flow
Friends
Order
Fitness
100% S
0.516129 0.322581
0.677419
0.322581
0.677419
11(2431)
0.700000
100% C
0.548387 0.322581
0.677419
0.322581
0.677419
10(2413)
0.666667
25% C
0.677419 0.322581
0.677419
0.322581
0.677419
10(2413)
0.700000
5% S
0.677419 0.322581
0.677419
0.838710
0.677419
10(2413)
0.700000
The Bad
Speed
Closest
Density
Flow
Friends
Order
Fitness
50% S
0.451613 0.935484
0.064516
0.677419
0.322581
21(4231)
0.600000
100% C
0.677419 0.322581
0.161290
0.838710
0.677419
10(2431)
0.533333
5% C
0.516129 0.903226
0.290323
0.225806
0.322581
29(1234)
0.666667
25% S
0.838710 0.870968
0.322581
0.677419
0.322581
19(4132)
0.633333
The Ugly
• Since multi-agents were spread throughout
the clump they influenced the other “dumb”
agents in ways which enabled them to get
out faster
• Clumps of evolving agents together form a
small pack which can increase exit speed by
not getting trapped behind other agents
• Usually got out under 6 seconds
Comments and Future
• Hard to debug and figure how multi-agents
respond
• Sheep herding (aren’t we all just sheep)
encouraging people exiting stadiums
• More complex environments/ distributions
• One final example/moral
Questions, Comments,
Cries of Joy?