Deep reinforcement learning for continuous-time self-triggered control with experimental evaluation

This paper develops a deterministic policy gradient (DPG) method for self-triggered control (STC) on continuous-time systems. To this end, we propose a reinforcement learning (RL) algorithm for STC. In general, control systems have high sampling rate (communication cost) to guarantee control performance. However, STC policy allows reducing the sampling rate while maintaining reasonable control performance. Moreover, it compresses the size of the historical data set, slowing down the data explosion. The STC-based RL demonstrates high data efficiency and the potential for industrial applications. A numerical simulation of a rotary inverted pendulum and the corresponding experiments are employed to show the effectiveness of the algorithm.