The tasks of behavior planning and decision-making learning in a dynamic environment are usually divided and considered separately in control systems for intelligent agents. A new unified hierarchical formulation of the problem of simultaneous learning and planning (SLAP) is proposed in the context of object-oriented reinforcement learning, and an architecture of a cognitive agent that solves this problem is described. A new algorithm for learning actions in a partially observed external environment is proposed using a reward signal, an object-oriented subject description of the states of the external environment, and dynamically updated action plans. The main properties and advantages of the proposed algorithm are considered, including the lack of a fixed cognitive cycle necessitating the separation of planning and learning subsystems in earlier algorithms and the ability to construct and update the model of interaction with the environment, thus increasing the learning efficiency. A theoretical justification of some provisions of this approach is given, a model example is proposed, and the principle of operation of a SLAP agent when driving an unmanned vehicle is demonstrated.
At Math-Net.Ru (available in 3 years after the publication): http://mi.mathnet.ru/eng/at/y2022/i6/p53
Panov A. I. Simultaneous Learning and Planning in a Hierarchical Control System for a Cognitive Agent // Autom Remote Control 83, 869–883 (2022).