Utilizing FORCE Learning to Model Adaptive Behavior

Abstract

Humans are able to learn from the environment and show a wide range of adaptive behaviors to solve the task at hand. They learn by trial and error. While reinforcement learning allows for artificial agents to learn via trial and error, they do so with algorithms that might not be the most biologically plausible. Recently, a new algorithm to train recurrent neural networks called FORCE learning has been proposed and this way of learning might be a lot more biologically plausible. We would like to research whether we can utilize this new algorithm to model adaptive behavior. Performance on a set of three toy problems was evaluated and it was shown that these agents were indeed able to learn to perform these tasks. Interestingly, this way of learning showed phenomena that are comparable to phenomena found in biological brains.