Inverse Reinforcement Learning for Optimal Consensus in Input‐Delay Multiagent Systems With Saturating Actuators

ABSTRACT

This paper considers the inverse reinforcement learning (IRL) problem of multi‐agent systems (MASs) in the presence of input delay and input saturation. The problem is formulated as Graphical Apprentice Games. To deal with the input delay, a model reduction method is employed to reconstruct the system state, and new performance indices are defined. Thus, the IRL problem of original MASs with input delay is transformed into the equivalent delay‐free counterpart. On this basis, a model‐based IRL algorithm is proposed, which recovers the expert's underlying performance functions via demonstrations. The algorithm consists of an optimal control (OC) learning stage and an inverse optimal control (IOC) update stage. Furthermore, when the agents' dynamics are completely unknown, a model‐free IRL algorithm is developed, which is implemented via neural networks (NNs). Numerical examples are provided to demonstrate the effectiveness of the proposed methods.