Multiagent fuzzy-neural control of a 3-link uniped

Abstract

Kgroo, a simulated 3-link folding legged uniped robot is presented and locomotion training of Kgroo with fuzzy-neural control is discussed. It is observed that for the uniped locomotion problem, global training of a fuzzy or neural controller is subject to failure. It is shown that, starting with a single jump example, a multiagent cerebellum model (MAC-J) can enable Kgroo to learn different jumps with a geometrical learning rate based on a learning-tuning-brainstorming theory. Technically, this work introduces effective means for decomposing the high-dimensional locomotion control problem into kernel spaces; theoretically, incremental learning and coordinated cerebellar agent discovery provide a natural explanation to certain explosive learning behaviors in human and animal locomotion control.