RULE ABSTRACTION
Don Miner
Multi-Agent Planning & Learning Lab
University of Maryland, Baltimore County
November 5, 2008
don1@umbc.edu
http://maple.cs.umbc.edu/~don/projects/SAF
SWARM INTELLIGENCE
• A sub-field of “Artificial Intelligence”
• Emergence - complex systems and patterns from simple interactions
• Properties:
• Collective Behavior
• Decentralized
• Self-organized
• Often is biologically-inspired
Borg Cube from Star Trek
EXAMPLE: BOIDS
Source: http://www.red3d.com/cwr/boids/
• “Boids” by Craig Reynolds in 1986
• Agents follow three rules:
• Separation Separation
• Alignment
• Cohesion Cohesion
• Result: flocking behavior Alignment
3
EXAMPLE: ANT COLONY OPTIMIZATION
Source: http://iridia.ulb.ac.be/~mdorigo/ACO/index.html
• “Optimization, Learning and Natural
Algorithms” by Marco Dorigo in 1992
• Ants forage for food and drop pheromones in
their path
• Ants follow the pheromones to food
• Over time, a short path is found
EXAMPLE: SWARM-BOTS
• A project headed by Marco Dorigo
• One of the first “Swarm Robotics” projects
• Many robots work independently
• Used as a platform for many applications
• Robots can form chains, collaborate on carrying
heavy objects, traverse obstacles, and more
SWARM-LEVEL PROPERTIES
• We (humans) sense abstract, swarm-level properties of swarms
• How to determine these properties from the available parameters is not clear
• Example “Form Circle”:
• Density (agents per pixel), radius, stability: swarm-level properties
• Center, Avoid, Circle Align: agent-level rules
6
SWARM APPLICATIONS
• Emergent behavior is created by many agents’ behavior
• Agent behavior is rooted in their rules
• Modifying the emergent behavior requires modifying the agents’ rules
• Wouldn’t it be more intuitive to modify emergent behavior at the swarm-level?
Behavior
Emergent Behavior
Adjust? Agent
Swarm
Avoid Center Align
7
EXAMPLE: “COHESIVENESS”
ReynoldsBoids(# agents, avoidance, center, alignment)
ReynoldsBoids(25, ?, ?, ?)
229, .05, .34 175, .22, 4.2
12.8 89.4
ReynoldsBoids(25, ?)
ReynoldsBoids(# agents, cohesiveness)
8
RULE ABSTRACTION
• Rule abstraction: the process of learning swarm-level properties
• Enables control of swarms in terms of the swarm-level properties
• Enables predictions of emergent behavior from the agent-level parameters
• Results: end-user control over swarms, swarm “planners”, richer applications
• How do we do this? We learn the correlations between swarm- and agent-
levels
Rule (r1)
Rule (r2)
Abstract a=
Rule (a) T(r1, r2, ... , rn)
Rule (rn)
9
“FORM CIRCLE”
• Agents form circles with three low-level rules:
• repulsion factor,
• attraction to center of mass,
• attraction to average distance from
center, of mass
• Abstract rule: Form circle of radius r
• Learned with equal interval sampling
and linear interpolation
10
VIDEO: “FORM CIRCLE”
11
“FORM CIRCLE” 250
Radius
• Varied two parameters for the 0
sample: number of agents 8000
tor
and avoidance factor 3
Numbe an
ce
F ac
r of Ag oid
ents Av
150
• Evenly spaced sampling 24
• Use interpolation to answer the questions: 3
50
75
- What radius will N agents 100
125
with avoidance A give?
- What N and A should we use
Number of Agents
150
to get a specific radius?
175
3
0
Number of Agents
200
Radius
24
150 Avoidance Factor 8000
250
24
150 8000 150 8000
Avoidance Factor 12
FUTURE WORK
• On-line learning/searing of satisfying constraints
(opposed to ahead of time sampling)
• Handling dynamic swarm behavior
(something learned one day may not be true the next)
• Abstraction hierarchies
(can we use abstract rules as low-level rules?)
• Control applications
(actually writing GUIs that can control and visualize swarms)
• Domains
(need more domains to be implemented to show my method works)
13
QUESTIONS?
Don Miner
Multi-Agent Planning & Learning Lab
University of Maryland, Baltimore County
don1@umbc.edu
http://maple.cs.umbc.edu/~don/projects/SAF