Learning to Control at Multiple Time Scales

In reinforcement learning the interaction between the agent and the environment generally takes place on a fixed time scale, which means that the control interval is set to a fixed time step. In order to determine a suitable fixed time scale one has to trade off accuracy in control against learning complexity. In this paper, we present an alternative approach that enables the agent to learn a control policy by using multiple time scales simultaneously. Instead of preselecting a fixed time scale, there are several time scales available during learning and the agent can select the appropriate time scale depending on the system state. The different time scales are multiples of a finest time scale which is denoted as the primitive time scale. Actions on a coarser time scale consist of several identical actions on the primitive time scale and are called multistep actions (MSAs). The special structure of these actions is efficiently exploited in our recently proposed MSA-Q-learning algorithm. In this paper, we use the MSAs to learn a control policy for a thermostat control problem. Our algorithm yields a fast and highly accurate control policy; in contrast, the standard Q-learning algorithms without MSAs fails to learn any useful control policy for this problem.