A Novel Adaptive Tropism Reward ADHDP Method with Robust Property

Abstract

According to the autonomous learning problem for the two-wheeled self-balancing robot, a novel adaptive tropism reward ADHDP with robust property was proposed, which can get the online adaptive tropism reward information. The whole learning system used a form of three networks, including action neural networks (ANN), adaptive tropism reward neural networks (ATRNN) and critic neural networks (CNN). The design of adaptive tropism reward neural networks took example from the learning mechanism of actor-critic structure. And through the primary binary reward signal, the continuous secondary reward signal can be got adaptively and become the basis of critic neural networks learning. Through the simulation in two-wheeled self-balancing robot, we can conclude that the proposed learning mechanism is effective and has a better progressive learning property. The optimal learning performance is got finally. Through the comparison of statistical experiment, it can be found that the proposed method has a certain anti-noise ability and the robust learning performance is better.