Safe reinforcement learning for cooperative tracking consensus problem of discrete-time multiagent systems with control barrier functions

The cooperative tracking consensus problem for discrete-time multiagent systems (MASs) with state and input constraints is addressed by a novel safe reinforcement learning (RL) algorithm. In MASs, all agents communicate through a directed communication graph. The algorithm is developed within the policy iteration (PI) framework, relying solely on measured data from the system trajectories in the environment. Additionally, the reward functions are enhanced by the integration of control barrier functions (CBFs) and input constraint functions. The iterative value function’s monotonically non-increasing feature guarantees that the safe set stays forward invariant inside the PI framework. Furthermore, theoretical proofs are provided to establish both safety and convergence. The algorithm’s design and implementation are elaborated through neural networks. Finally, experiment results validate the efficacy and safety of the proposed algorithm.