A New Hybrid Learning Algorithm for Drifting Environments

Abstract

An adaptive algorithm for dnfting environments is proposed and tested in simulated environments. Two simple but powerful problem solving technologies - Neural Networks and Genetic Algorithms with Online Learning, help the artificially intelligent agents to adapt to a changing environment. Neural networks and genetic algorithms are combined to evolve weights, architecture, and learning rules for the generation of efficient networks. Online learning helps these networks to capture the dynamics of a changing environment efficiently. Supervised learning 1s achieved using a variation of regular backpropagation that works on dynamic random networks.

Our algorithm proposes two types of online learning, namely local online learning which requires a pre-defined training set and global online learning which does not It is shown that both types of online learning improve the performance of networks to capture subtleties of the varying environments.

The algorithm's efficiency is demonstrated using a mine sweeper application. Different learning technologies have been compared. The results establish that online learning within the evolutionary process is the most significant factor for adaptation and 1s far superior to evolutionary algorithms alone. The evolution and learning work in a co-operating fashion to produce excellent results in short time. Offline learning is observed to increase the average fitness of whole population. It is also demonstrated that online learning is self sufficient and can achieve results without any pre-training stage. When mine sweepers are able to learn online, their performance in the drifting environment is significantly improved.