本研究主要是針對傳統小腦模型控制器(Cerebellar Model
Articulation Controller, CMAC)的學習控制理論，提出一個新的
「植基於模糊推論之小腦模型控制器學習控制法則」(Fuzzy-
Inference-Based Cerebellar Model Articulation Controller;
FBCMAC)，以改進傳統小腦模型控制器的學習精度及學習速度。傳統
小腦模型控制器學習速率是設計為定值，取大的學習速率則學習速
度可較快但可能引起不穩定，反之取小的學習速率則學習速度有較
穩定的特性、較精確的學習，但學習遠度會變慢，所以要獲得較好
的學習結果就必須找到一個合適的學習速率，但可想而知，要找到
一個適合的學習速率是相當耗時的事情，本研究主要設計理念是針
對傳統小腦模型控制器的學習速率為定值的設計方式，提出新的學
習法則，利用模糊推論依據系統狀況找出其較佳之學習速率，達到
動態調整學習速率，使整個學習加速、學習精度改進，本研究提出
三種新的植基於模糊推論之小腦模型控制器演算法，週期式、狀態
式以及切換式植基於模糊推論之小腦模型控制器演算法，並且探討
新小腦模型控制器學習法則在學習控制上優於傳統小腦模型控制器
的主要特性。

The paper proposed a new fuzzy-inference-based Cerebellar Model Articulation Controller (FBCMAC) to improve the learning convergent error and learning convergent speed of conventional Cerebellar Model Articulation Controller (CMAC). The learning rate of conventional CMAC is constant value. If the learning rate is a bigger value then the convergent speed is more quick, but it maybe result in unstable phenomenon; contradictorily, a smaller value of learning rate has a more stable learning property and more precise learning result, but the learning speed is slower. If we want to get better learning result, we must search a suitable learning rate. However, it must spend amount of time to find a suitable learning rate. The main goal of the paper is to improve the drawback of conventional CMAC that learning rate is constant value. The paper proposed a new learning rule that utilized the fuzzy inference to find a better learning rate. It can adjust dynamically the learning rate. Let the entire learning performance speed up and improve the learning precision. In the paper, we proposed three different learning algorithms that are based on fuzzy inference to get dynamically the learning rate. They are respectively cycle style, state style and switching style. We discuss the properties of new learning algorithm, and guarantee the new learning algorithm is superior to the conventional learning algorithm.