Multimodal Control in Uncertain Environments using Reinforcement Learning and Gaussian Processes

The control of complex systems can be done decomposing the control task into a sequence of control modes, or modes for short. Each mode implements a parameterized feedback law until a termination condition is activated in response to the occurrence of an exogenous/endogenous event, which indicates that the execution mode must end. This paper presents a novel approach to find an optimal switching policy to solve a control problem by optimizing some measure of cost/benefit. An optimal policy implements an optimal multimodal control program, consisting in a sequence of control modes. The proposal includes the development of an algorithm based on the idea of dynamic programming integrating Gaussian processes and Bayesian active learning. In addition, an efficient use of the data to improve the exploration of the continuous state spaces solutions can be achieved through this approach. A representative case study is discussed and analyzed to demonstrate the performance of the proposed algorithm