Deep learning of adaptive control systems based on a logical-probabilistic approach

The problem of automatic selection of subgoals is currently one of the most relevant in adaptive control problems, in particular, in Reinforcement Learning. This paper proposes a logical-probabilistic approach to the construction of adaptive learning control systems capable of detecting deep implicit subgoals. The approach uses the ideas of the neurophysiological Theory of functional systems to organize the control scheme, and logical-probabilistic methods of machine learning to train the rules of the system and identify subgoals. The efficiency of the proposed approach is demonstrated by an example of solving a three-stage foraging problem containing two nested implicit subgoals.