Preventing Overlearning in CMAC by Using a Short-Term Memory

Abstract

This paper proposes improvements to the cerebellar model articulation controller (CMAC). Often, the CMAC plays the role of a fuzzy nonlinear approximator in direct adaptive control. However, in applications where the inputs oscillate the CMAC exhibits more drift in the adaptive parameters (overlearning) compared to other types of fuzzy approximators and neural networks and is thus more likely to cause bursting. This is due to the local nature of CMAC fuzzy membership functions: even a small oscillation of the trajectory across the origin and through different membership functions results in repeated positive or negative updates for each adaptive parameter. The solution proposed in this paper uses a short-term memory: a membership function remains activated during the indexing of subsequent CMAC cells. This ensures a small oscillation will occur entirely within membership functions rather than between them. The method uses Gaussian radial basis functions. Simulation results with an underdamped flexible joint demonstrate the improved performance and stability characteristics.