Enhancing a Model-Free Adaptive Controller through Evolutionary Computation

摘要

Many robotic systems experience fluctuating dynamics during their lifetime. Variations can be attributed in part to material degradation and decay of mechanical hardware. One approach to mitigating these problems is to utilize an adaptive controller. For example, in model-free adaptive control (MFAC) a controller learns how to drive a system by continually updating link weights of an artificial neural network (ANN). However, determining the optimal control parameters for MFAC, including the structure of the underlying ANN, is a challenging process. In this paper we investigate how to enhance the online adaptability of MFAC-based systems through computational evolution. We apply the proposed methods to a simulated robotic fish propelled by a flexible caudal fin. Results demonstrate that the robot is able to effectively respond to changing fin characteristics and varying control signals when using an evolved MFAC controller. Notably, the system is able to adapt to characteristics not encountered during evolution. The proposed technique is general and can be applied to improve the adaptability of other cyber-physical systems.