Direct heuristic dynamic programming based on an improved PID neural network

As an online learning algorithm of approximate dynamic programming (ADP), direct heuristic dynamic programming (DHDP) has demonstrated its applicability to large state and control problems. However, there still lacks of a systemic approach to initialize the network weights for DHDP. In this paper, an improved PID-neural network (IPIDNN) configuration is proposed and applied to the critic and action networks of DHDP, which is flexible and easy to expand. Because of incorporating an inherent PID control structure, it is easy to use a well-designed PID controller to guide the initial weighs choosing for the action network. Based on this framework, a novel initializing approach is suggested based on a PID controller, such that the DHDP learning process starts from a good enough initial state. Simulations are carried on a cart-pole system to validate the effectiveness of the IPIDNN-based DHDP and the proposed initializing approach.