Safe deep reinforcement learning-based adaptive control for USV interception mission

This paper aims to develop a safe learning scheme of the USV interception mission. A safe Lyapunov boundary deep deterministic policy gradient (SLDDPG) algorithm is presented for the USV interception mission. The uniformly ultimate bounded (UUB) stability of control systems is analyzed under finite safety constraints. A single neuron proportional adaptive control (SNPAC) is applied to pre-train the deep policy network for speeding up the training process. The proposed method is evaluated by a series of simulations of the USVs interception and tracking mission. Compared with the existing results, our method can fast converge to the feasible solution subject to safety constraints and demonstrate a high performance in stability and safety by virtual-reality experiments.

Automated summary

This paper presents a safe learning scheme for the USV interception mission using a safe Lyapunov boundary deep deterministic policy gradient algorithm. By applying a single neuron proportional adaptive control for pre-training, the proposed method quickly converges to feasible solutions subject to safety constraints and demonstrates high performance in stability and safety.