Multi-agent system approach to react to sudden environmental changes

Abstract

Many processes experience abrupt changes in their dynamics. This causes problems for some prediction algorithms which assume that the dynamics of the sequence to be predicted are constant, or at least only change slowly over time. In this paper the problem of predicting sequences with sudden changes in dynamics is considered. For a model of multivariate Gaussian data we derive expected generalization error of standard linear Fisher classifier in situation where after unexpected task change, the classification algorithm learns on a mixture of old and new data. We show both analytically and by an experiment that optimal length of learning sequence depends on complexity of the task, input dimensionality, on the power and periodicity of the changes. The proposed solution is to consider a collection of agents, in this case non-linear single layer perceptrons (agents), trained by a memetic like learning algorithm. The most successful agents are voting for predictions. A grouped structure of the agent population assists in obtaining favorable diversity in the agent population. Efficiency of socially organized evolving multi-agent system is demonstrated on an artificial problem. © Springer-Verlag Berlin Heidelberg 2007.