Ventilation control learning with FACL

Fuzzy actor-critic learning (FACL) is a reinforcement learning method that tunes fuzzy controllers (FC). Based only on reinforcement signals, such as rewards and punishments, that describe the control task, FACL qualifies FC's actions to approximate optimal policies. One of the most important user step is to define good reinforcement functions. In this article, we introduce fuzzy reinforcement functions (FRF) to describe the task in such a way that the frontiers between success and failure states become smooth. This new type of reinforcement function brings more informations than the classical one, allowing a higher learning speed. We apply these FRFs with FACL on an industrial task that consists in controlling a building atmosphere.