Accelerated Variant of Reinforcement Learning Algorithms for Light Control with Non-Stationary User Behaviour

In the context of smart building energy management, we address in this work the problem of controlling
light so as to minimise energy usage of the building while maintaining the satisfaction of the user regarding
comfort using a stateless Reinforcement Learning approach. We consider that the user can freely interact with
the building and changes the intensity of the light according to his comfort. Moreover, we consider that the
behaviour of the user depends not only on present conditions but also on past behaviour of the control system.
In this setting, we use the pursuit algorithm to control the signal and investigate the impact of the discretization
of the action space on the convergence speed of the algorithm and the quality of the policies learned by the
agent. We propose ways to accelerate convergence speed by varying the maximal duration of the actions while
maintaining the quality of the policies and compare different solutions to achieve it.