Particle Swarm Optimization and Social Interaction Between Agents Kenneth Lee TJHSST 2008
Overview Overview of PSO Background Research Project Goals Types of Social Interactions Project State/Results Conclusion
Overview Of PSO Originally designed to model birds Overtime became more analogous to a swarming animal (bees) Search for Global Optima Infinite search spaces
Overview Of PSO “Particles” (vectors) Random Position Random Velocity Influences on Velocity Cognitive Influence Social Influence Convergence(?)
Particle Swarm Optimization for k = 1 to number of particles n do if (fitness(k) < fitness_lbest(k)) lbest(k) = pos(k) endif end do for k = 1 to number of particles n do social(k) enddo for k = 1 to number of particles n do for I = 1 to number of dimensions d do R1 = randomNumber R2 = randomNumber V[k][I] = w * (C1 * R1 * (pos-lbest) + C2 * R2 * (pos-gbest)) X[k][I] = pos + V[k][I] enddo Determining lbest Social Interaction Adjusting Position
Importance of Social Interaction Influences Velocity V = ??? X’ = X + V Encourages Exploration Through Social Interaction, Particles are able to communicate information and extrapolate data about the objective function.
Social Interactions Variance of k value (# of neighbors) Through research k values between 3-5 seem to work best Topology? Cliques Random Share/Follow
Project State 5 Interactions NIPS SIPS RIPS FIPS DIPS 3 Benchmark Functions Rastrigin, Six Camel Hump, Sphere
NIPS (Non-Informed Particle Swarm)
SIPS (Singly-Informed Particle Swarm)
RIPS (Ring Informed Particle Swarm)
FIPS (Fully Informed Particle Swarm)
DIPS (Dynamically Informed Particle Swarm)
Conclusions DIPS seems to perform best Time only DIPS and RIPS have 100% success rate FIPS converges fastest