Particle Swarm Social Model for Group Social Learning in an Adaptive Environment

1
Particle Swarm Social Model for Group Social
Learning in an Adaptive Environment
Xiaohui Cui, Laura L. Pullum, Jim
Treadwell, Robert M. Patton, and Thomas E.
Potok
Computational Sciences and Engineering
Division Oak Ridge, TN 37831 865-576-9654 cuix_at_orn
l.gov
3333 Pilot Knob Road Eagan, MN
55121 651-456-3247 Laura.L.Pullum_at_lmco.com
2
Research Overview

• Integrate particle swarm algorithm, social
  knowledge adaptation and multi-agent approaches
  for modeling the social learning of
  self-organized groups and their collective
  searching behavior in an adaptive environment.
• Apply the particle swarm metaphor as a model of
  social learning for a dynamic environment.
  Provides an agent-based simulation platform for
  understanding knowledge discovery and strategic
  search in human self-organized social groups.
• Investigate the factors that affect the global
  performance of the whole social community through
  social learning.

3
Particle Swarm Social Learning Model in Adaptive
Environment

• Social Learning
• Learning by observing models and noting the
  reward contingencies
• Adaptive Environment
• Models in the environment change in each time
  step. The change can be linear or random.
• Highly rewarded models dynamically change in
  every time-step. The observed highly rewarded
  models by learner in time t1 may not be
  rewarded models in time t2.
• Change patterns of the environment are influenced
  by the collective behavior of the learner groups
  when this collective behavior is effective enough
  to alter the environment

• Particle Swarm algorithm
• Developed in 1995 by James Kennedy and Russ
  Eberhart
• Inspired by social behavior of bird flocking
• Applies the concept of social interaction to
  problem solving
• Been applied to a wide variety of search and
  optimization problems
• Particle swarm social learning model
• Every particle is considered as a human group
• Particles interact with each other
• The group can learn skills and behaviors by
  observations
• Particles are more likely to imitate models whose
  behavior is rewarded
• Particle also has a memory of its behavior
  history (e.g., people can learn from their own
  experiences)

Personal Cognition
Social Adaptive Learning
4
Experiment Results

• Adaptive Environment
• A two dimensional DF1 equation (1) is used to
  produce the dynamic environment. The environment
  change rate is controlled through the logic
  equation (2). The adaptive mechanism of the
  environment is represented by equation (3). The
  fitness value of the solution gradually decreases
  when an increasing number of the group members
  search for problem solution in the neighbor area.

(1)
(2)
(3)
Figure 1 The sample landscape environment
Figure 2 The step size value map generated by
equation (2) with different A value

• Experiment Setup
• With following experiment setup, Figure 3
  illustrates the initial simulation environment
  with 20 agent groups.

Figure 3 The initial Environment Agent Group

• Experiment Results
• Results from the simulation have shown that
  effective communication is not a necessary
  requirement for self organized groups to attain
  higher profit in an adaptive environment

(a)
(b)
(b)
(a)
Figure 5 The comparison of the average fitness
values of (a) each simulation iteration for group
scenario a and b (b) whole simulation for
different agent group scenarios
Figure 4 The collective searching results for
scenario (a) one group with 400 agents and
scenario (b) twenty groups, 20 agents per group
5
Verification and Validation

• Verification requires rigorous standardized
  test problems, benchmarks
• Benchmark problems moving parabola problem,
  moving peaks benchmark function, DF1 DEFEAT
  Test Environment
• Formal Methods used for NASA ANTS verification

• Validation
• Compare agent-based simulation/system (ABS/S)
  output with real phenomenon
• Compare ABS/S results with math model results
• Dock with other simulations of same phenomenon

Data Validation Sources Types
6
Conclusions Future Direction

• The dynamics of the real world are influenced by
  the collective actions of social groups
• The changes of the real world impact the social
  groups actions and structure
• The Particle swarm social learning model is
  developed to simulate the complex interactions
  and the collective searching of the
  self-organized groups in an adaptive environment
• Next steps
• More sophisticated models
• Reaction, perception of environment
• Access to more real data
• Increased validation of models
• Extend application to related domains