H_∞ reinforcement learning control of robot manipulators using fuzzy wavelet networks

摘要

In this paper, an H_∞ reinforcement learning controller based on a fuzzy wavelet network (FWN) is proposed to perform a position-tracking task for a robot manipulator. The proposed controller adopts the actor-critic reinforcement learning control scheme. The primary reinforcement is generated by a performance measurement unit. The learning unit of the controller consists of an associative search network (ASN) and an adaptive critic network (ACN). The ASN is employed to approximate unknown nonlinear functions in the robot dynamics and the ACN is utilized to construct a more informative signal than the primary reinforcement alone to tune the ASN. Since the FWN can provide accurate function approximation, both the ASN and ACN are implemented by the FWN. In addition, the proposed controller requires no prior knowledge about the dynamics of the robot manipulators and no off-line learning phase. Moreover, by employing the H_∞ control theory, it is possible to attenuate the effects of the approximation errors of the FWNs and external disturbances to a prescribed level. In contrast to the general H_∞ problem, only simple equations, rather than Riccati equations, should be solved. Computer simulations on a SCARA robot with 3 degrees-of-freedom confirm the effectiveness of the FWN-based controller with H_∞ stabilization.