An Alternative CMAC Trained with Genetic Algorithms to Solve an Instable Control Process
Yun‐Kung Chung, Shing-Chung Hu · 2005
Since CMAC (Cerebellar Model Articulation Controller), which mimics the recognition process of human cerebellum, is an artificial neural network (ANN) characterized by the fast learning or high convergence, it is easy to be programmed for online learning and real-time control. Diversified applications have exhibited the learning and solution abilities of CMAC; nevertheless, CMAC has an unstable learning performance in certain experimental cases presented by Chen and Chang [4, 5, 6]. To overcome the above CMAC learning instability, a GA (Genetic Algorithm) is used as an alternative approach to train CMAC in this paper. GA is an optimization technique that mimics genetic changes in the life evolutionary process. It takes advantage of the multiple starting points for searching an optimal solution within a relatively short computational time. By comparing the standard CMAC and the GA-based CMAC (GACMAC), the usefulness of the GACMAC for overcoming the CMAC learning instability was verified by experimenting the Chan and Chang's time series control problem. More general cases will be needed to study for the confirmation of the alternative CMAC training performance in the near future.