Fitted Q-Function Control Methodology Based on Takagi–Sugeno Systems

Abstract

This paper presents a combined identification/ Q-function fitting methodology that involves identification of a Takagi-Sugeno model, computation of (sub)optimal controllers from linear matrix inequalities (LMIs), and subsequent data-based fitting of the Q-function via monotonic optimization. The LMI-based initialization provides a conservative solution, but it is a sensible starting point to avoid convergence/local-minima issues in raw data-based fitted Q-iteration or Bellman residual minimization. An inverted-pendulum experimental case study illustrates the approach.