A novel method for optimal consensus control of multi-agent systems (MASs) based on adaptive dynamic programming (ADP) is developed in this paper. Unlike neural networks (NNs) that require manually designed features in value function approximation and may effect the approximation quality. Kernel-based methods are adopted to approximate value functions without predefining the model structure. Moreover, to handle the challenge of unknown or complex system dynamics, a local action-value function is defined and kernel-based methods are used to approximate the local action-value function. Thus, an ...