Barrier function based safe reinforcement learning for multi-agent systems

This paper intends to design a safety optimized controller that can guarantee collision avoidance and maintaining communication among multi-agent systems (MAS) while minimizing some performance. Different from the existing works on safety optimization problems using quadratic program (QP) method, a new class of Lyapunov-like barrier functions (BFs) is introduced and integrated into the performance indices to guarantee safety. This can transform the original constrained optimal control problem into an unconstrained one. Furthermore, the vanishing viscosity method is introduced to construct a general value function which eliminates the effect of nonsmooth caused by the input constraints on the solution of the modified Hamilton-Jacobi-Bellman (HJB) equation. To solve the HJB equation, an improved Actor-Critic (A-C) neural networks (NNs) algorithm is developed to find a smooth approximation safety-optimized controller. Finally, a simulation is performed to demonstrate the effectiveness of the proposed approach.