Towards Incremental Social Learning in Optimization and Multiagent Systems

by Marco A. Montes de Oca and Thomas Stützle March 2008

Submitted to the Evolutionary Computation and Multi-Agent Systems and Simulation Workshop (ECoMASS-2008).

This page contains all supplementary information that, for the sake of conciseness, was not included in the paper.

Table of Contents
  1. Paper Abstract
  2. Incremental Social Learning in Optimization
  3. Incremental Social Learning in Multiagent Systems

Paper Abstract

Social learning is a mechanism that allows individuals to acquire knowledge from others without incurring the costs of acquiring it individually. Individuals that learn socially can thus spend their time and energy exploiting their knowledge or learning new things. In this paper, we adapt these ideas for their application to both optimization and multiagent learning. The approach consists of a growing population of agents that learn socially as they become part of the main population. We find that learning socially in an incremental way can speed up the optimization and learning processes, as well as improve the quality of the solutions and strategies found.

Keywords: Social Learning, Particle Swarm Optimization, Multiagent systems.



Incremental Social Learning in Optimization

The following figures show the median solution quality improvement over time. Different population sizes are used for the traditional particle swarm optimization algorithm and different stagnation thresholds are used for the incremental social learning approach.

Ackley function - 100 dimensions
Best particle as modelRandom particle as model
Fully connected topology Ring topologyFully connected topology Ring topology
Ackley - FC Ackley- R Ackley - FC Ackley- R
Rastrigin function - 100 dimensions
Best particle as modelRandom particle as model
Fully connected topology Ring topologyFully connected topology Ring topology
Ackley - FC Ackley- R Ackley - FC Ackley- R
Rosenbrock function - 100 dimensions
Best particle as modelRandom particle as model
Fully connected topology Ring topologyFully connected topology Ring topology
Ackley - FC Ackley- R Ackley - FC Ackley- R
Schwefel function - 100 dimensions
Best particle as modelRandom particle as model
Fully connected topology Ring topologyFully connected topology Ring topology
Ackley - FC Ackley- R Ackley - FC Ackley- R
Step function - 100 dimensions
Best particle as modelRandom particle as model
Fully connected topology Ring topologyFully connected topology Ring topology
Ackley - FC Ackley- R Ackley - FC Ackley- R

Incremental Social Learning in Multiagent Systems

The following figures show the median collective reward over time. Different agent addition schedules are used for the implementation of the incremental social learning framework.

Learning Scenarios
Cliff environmentClustered environment
Ackley - FC Ackley- R