We develop reinforcement learning techniques for energy efficient operation of buildings and systems without the need for mathematical models. Despite the many advantages of RL for application in the built environment, many challenges remain, and are explored in our research.
(1) As RL is a relatively new and emerging field, there are no off-the shelf implementations available that researchers could experiment with, in contrast to supervised or unsupervised methods.
(2) RL approaches have generally long learning times compared to model based approaches. This is typically the case for systems starting with zero prior knowledge. In buildings, however, a substantial prior and expert knowledge exists that can be leveraged to initialize the controllers.
(3) In the built environment, we have many potential learning agents, which naturally constitute a multi agent system. Multi-agent RL has been recognized as the most suitable approach to tackle large scale complex real-world problems. However, the theoretical field is still in its infancy, and most available results are for two agents. The presence of multiple agents violates the stationary environment underlying most single agents learning approaches. Here, our research in the built environment can define clear goals and metrics by which the learning agents are to be evaluated to guide the theoretical developments.