Actor-Critic Off-Policy Learning for Optimal Control of Multiple-Model Discrete-Time Systems

In this paper, motivated by human neurocognitive experiments, a model-free off-policy reinforcement learning algorithm is developed to solve the optimal tracking control of multiple-model linear discrete-time systems. First, an adaptive self-organizing map neural network is used to determine the system behavior from measured data and to assign a responsibility signal to each of system possible behaviors. A new model is added if a sudden change of system behavior is detected from the measured data and the behavior has not been previously detected. A value function is represented by partially weighted value functions. Then, the off-policy iteration algorithm is generalized to multiple-model learning to find a solution without any knowledge about the system dynamics or reference trajectory dynamics. The off-policy approach helps to increase data efficiency and speed of tuning since a stream of experiences obtained from executing a behavior policy is reused to update several value functions corresponding to different learning policies sequentially. Two numerical examples serve as a demonstration of the off-policy algorithm performance.