[2E6-GS-8] Robot and real worlds:

[2E6-GS-8-03] Event-Triggered Reinforcement Learning for Optimization of Online Control Systems

〇Hayato Chujo¹, Sachiyo Arai¹ (1. Chiba University)

Keywords:Event-triggered reinforcement learning, Semi-Markov decision process, HVAC control

In recent years, research on optimization of control systems using online reinforcement learning, which simultaneously learns measures and controls with the measures, has been progressing. We focus on event-driven reinforcement learning as an approach to optimize both control operations and time intervals. Compared with time-driven reinforcement learning, which performs control operations at fixed time intervals, event-driven reinforcement learning can solve the problems of instability caused by unnecessary control operations and increased control cost. However, the performance of event-driven reinforcement learning tends to deteriorate in the early stages of learning due to the effect of initial settings, which is a cause of instability in control using online reinforcement learning.Therefore, we propose a combined time-driven and event-driven reinforcement learning method to improve the performance of event-driven reinforcement learning in the early stages of learning. We also evaluate the performance of the proposed method by conducting computer experiments assuming the control of a heater.