CREATING AN AGENT BASED ON HIERARCHICAL TEMPORAL MEMORY MODEL

Kanonirs Georgijs

Modern reinforcement learning methods have a number of limitations imposed by the used artificial neural networks paradigm with a point neuron model. The use of the latest achievements of neuroscience within a new theory of intelligence called The Thousand Brains Theory of Intelligence, as well as the application of the machine learning model called Hierarchical Temporal Memory (HTM), which implements some aspects of this theory, have the potential both to develop already established reinforcement learning methods, and to create new approaches for solving this problem. The purpose of this work is to propose and analyze a conceptual idea for creating an agent based on this model, allowing this task to be done in the most natural way, i.e. based on its underlying structural and functional principles, and without specific requirements for a reward signal.

Keywords:biologically-plausible machine learning methods, reinforcement learning, hierarchical temporal memory