Rule-Based Sequence Learning Extension for Animats

Abstract

This thesis introduces a rule-based, sequence learning model. It proposes that parts of this model could be used as a independent extension to other machine learning models, animats specifically. The model uses Q-learning and state space search to generalize which are equivalent. This allows reducing the input state space to train faster and better draw conclusions about the features in the dataset at large. This knowledge can then be used to calculate the best action for the given sequence. The model is implemented in order to evaluate its capabilities. The model is evaluated primarily on the domains of simple arithmetic, Boolean logic, and simple English grammar and then compared to the performance of a Recurrent Neural Network using Long-Short Term Memory-units.