Evolving Echo State Networks for Minimally Cognitive Unsupervised Learning Tasks

Abstract

This thesis investigates the use of Echo State Networks (ESNs) in unsupervised learning environments, byemploying Evolutionary Algorithms (EAs) to evolve ESNs to control an agent that performs a novel, minimally-cognitive learning task. The task employed in this thesis is a modified version of the classic video game Frogger. ESNs are investigated since they promise to combine the temporal abilities seen in other Recurrent Neural Networks (RNNs) with a straightforward method to train the network. However, previous work employing ESNs in unsupervised environments is lacking.The evolved ESNs are compared to feed-forward Artificial Neural Networks (ANNs) as well as ESNs trainedwith regular supervised learning, in a comparative performance measure to find out which method is bestsuited to control Frogger.The results from this thesis show that not only do ESNs work well with EAs, but they surpass traditional feed-forward ANNs on the Frogger task. Additionally, it is shown that for the Frogger task, evolved ESNs also outperform ESNs trained with supervised learning. The results from this work should serve as an encouragement to use ESNs for more tasks in the future, due to their competitive performance and ease of setup.