Online tuning of a supervisory fuzzy controller for low-energy building system using reinforcement learning

摘要

This paper proposes a model-free method using reinforcement learning scheme to tune a supervisory controller for a low-energy building system online. The training time and computational demands are reduced by basing the supervisor on sets of fuzzy rules generated by off-line optimisation and by learning the optimal values of only one parameter, which selects the most appropriate set of rules. By carefully choosing the tuning targets, discretizing the state space, parameterizing the fuzzy rule base, using fuzzy trace-back, the proposed method can complete the training process in one season.