THE STUDY OF CMAC WITH ADAPTIVE QUANTIZATION

• Main Authors: Chih-Cheng Chang, 張志正
• Other Authors: Hung-Ching Lu
• Format: Others
• Language: en_US
• Published: 2002
• Online Access: http://ndltd.ncl.edu.tw/handle/81580558168254043517

碩士 === 大同大學 === 電機工程研究所 === 90 === Cerebellar Model Arithmetic Controller (CMAC) is a kind of neural networks that imitate the human cerebellum and has attractive properties of learning convergence and speed. However, we often have to make our choice between accuracy and memory sizes when using CMAC. Therefore, quantization plays important role in construction of CMAC. In this thesis, we will focus on the quantization problem of CMAC. By using equal-sized quantization, CMAC cannot use finite knots to represent the variation of the reference signal efficiently. Therefore, an adaptive quantization rule is proposed to solve these problems. After adaptively quantizing, we can effectively sample input space and have better generalization property. At last, the simulation results will verify the proposed scheme.