Abstract:
The Cerebellar Model Articulation Controller (CMAC) type of neural network contains local basis-function domains defined by multiple offset arrays of hypercubes; it shows...Show MoreMetadata
Abstract:
The Cerebellar Model Articulation Controller (CMAC) type of neural network contains local basis-function domains defined by multiple offset arrays of hypercubes; it shows particular adeptness for control systems due to fast adaptation and the ability to handle many inputs. However, the weights in the CMAC tend to drift (i.e. adaptive-parameter drift or overlearning) due to the input oscillating between distinct local basis functions in each array. Robust weight update methods like e-modification often sacrifice performance for stability in the case of CMAC. Contrast this to radial basis function networks and multilayer perceptrons, where inputs always remain within global basis functions and weight drift is not particularly difficult to prevent. This paper proposes overlayering the basis functions in each CMAC array, by keeping a basis function activated even after its associated CMAC array cell is no longer indexed by the input. In this way, oscillations occur entirely within basis functions in each CMAC array. Simulations with a two-link, flexible-joint robot demonstrate that the proposed method avoids the trade-off between performance and stability found with the original CMAC.
Date of Conference: 27 November 2017 - 01 December 2017
Date Added to IEEE Xplore: 05 February 2018
ISBN Information: