INCREASING THE EFFICIENCY OF AGENT TRAINING BASED ON THE HIERARCHICAL TEMPORAL MEMORY MODEL

Kanonir Georgy, Filchenkov Andrey Aleksandrovich

Modern methods of reinforcement learning have a number of limitations imposed by the used paradigm of artificial neural networks with a point model of a neuron. The use of the "hierarchical temporal memory" (HTM) model has the potential both for the development of already established training methods and for the creation of new ones. The aim of this paper is to propose a new design of a spatial-temporal memory unit that allows an agent based on the HTM model to take into account a temporal context of limited length and, due to this, to increase the efficiency of its learning when solving problems in which the actual reward received depends on a temporal context of a size smaller than the maximum length of the sequences of observations and actions considered within the framework of the problem being solved.

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