Improvement of LMI controllers of Takagi-Sugeno models via Q-learning*

This paper presents a preliminary attempt to bridge the conservative (shape-independent) results from guaranteed-cost LMIs and the reinforcement learning setups which learn optimal controllers from data. In this sense, the proposed approach uses an initialization based on the LMI solution and proposes an approximation of the Q-function using polynomials of the membership functions in Takagi-Sugeno models. The resulting controller is shape-dependent, that is, uses the knowledge of membership functions and data to clearly improve LMI solutions.