Radial basis function networks for adaptive critic learning

An adaptive critic learning (ACL) structure typically consists of two main portions: the critic (evaluation) module and the action (control) module. The critic module learns how to evaluate the situation while the action module learns the control/decision-making skill. In this paper, radial basis function networks (RBFNs) are proposed for implementing these two learning modules. Results show that the RBFN-based ACL has a good learning speed. Using RBFNs, the ACL will have a better capability for solving larger size problems. While RBFs are differentiable, they are suitable for the action dependent critic (ADC) scheme, which requires the derivatives of the critic with respect to actions. The ADC is a more powerful learning scheme modified from ACL