Real-time optimal control for three-tank level system via improved ADDHP method

This paper presents a novel real-time optimal neural controller which is based on improved Action-Depended Dual Heuristic Dynamic Programming (ADDHP) method, including its schematic diagram, the training algorithms and its implementation steps. This method requires neither an explicit model of the controlled plant nor the indispensable system performance index `J´ which is explicitly defined in the classical optimal control theory. Compared with the traditional methods of Dual Heuristic Dynamic Programming (DHP) and ADDHP, the improved ADDHP method only uses the states of present and previous time step to calculate the derivative of the performance function, avoiding to predict the states of next time step, so the model network can be omitted. It makes the configuration of this method become more simple and more suitable for real-time application for complex nonlinear system or processes. An real-time control example is given, and the experimental results show that this ADDHP-base optimal neural controller has advantages in fast response, anti-jamming capability and robustness.