CMAC類神經網路控制系統及CMAC晶片實現

Abstract

本論文的目的主要可以分為系統分析以及硬體設計兩方面。在系統分析方面，本論文研究CMAC類神經網路控制系統的控制行為並分析系統的特性。這裡使用Runge-Kutta法來模擬受控系統的輸出/輸入特性。CMAC控制系統不需要關於機械臂的任何資訊，而且可以應付很大的負載變化。當輸入向量緩慢的移動的時候，CMAC控制器會產生龐大的積分動作。但是當輸入向量快速變化的時候，CMAC控制器也能有效的產生“遺忘”效應。論文中也證明CMAC控制系統能夠在目標位置收斂。在硬體設計方面，本論文規劃了CMAC類神經網路控制晶片的架構以及用硬體描述語言VHDL做更進一步的實現。在架構設計上，包括引用權重直接定址方式以心縮式陣列(Systolic Array)的結構來實現CMAC類神經網路中的映射關係，以及使用線性回饋暫存器(Linear Feedback Shift Register)所具備產生偽亂數的功能來簡化並完成雜湊編碼所能達到的目標等。最後，本論文提供此CMAC類神經網路控制晶片平行處理的解決方案，藉以增加晶片內部單元平行處理的套數並做適當的排程來大幅提升不同應用所可能需求的更高處理速度。

There are two main objective of this paper, system analysis and hardware design. In the aspect of system analysis, the behavior of the CMAC control system has been studied and stability of the system has been analyzed in this thesis. The Runge-Kutta method is adapted to find out the characteristic between output and input of plant. However, the CMAC requires no information about the robot, and can deal with large variations in load. The CMAC produces enormous integration action when the input vector moves slowly in the space, but it can also forget efficiently when the input vector moves fast in the space. It is shown that the CMAC control system can converge into the target position. In the aspect of hardware design, on the architecture design and further implementation in Hardware Description Language VHDL have been focused in this thesis. On architecture design, the systolic array structure in the proposed Direct Weight Address Mapping method is adapted to implement the mapping relationship in the CMAC neural network. And the hash coding function is simplified and achieved by generating pseudo-random function in the Linear Feedback Shift Register. At last but not least, the solution of parallel processing in the CMAC neural network chip is provided, for higher processing speed in accordance to different application.