結合BMF模糊類神經網路與實數基因演算於系統鑑別上的應用

Abstract

本文提出使用實數型基因演算法，找出最佳模糊類神經網路的權重值及BMF（Bspline membership function）控制點的方法。使用傳統模糊類神經網路透過梯度下降法學習，在學習過程中可能會產生落入區域極值的現象，在本文中使用BMP模糊類神經網路，藉著B-spline函數區間調整的特性，使系統作細微的調整。我們隨機建立初始的模糊評估法則，藉由循序搜尋單參數交配的實數基因演算法作為學習法則，並且由建立完整的自動學習機構來學習。我們使用實數基因演算法來解決二進制基因演算法在演化過程中，編碼解碼所造成複雜的運算。最後我們將它應用在離線調整，得到不錯的結果。

The main propose of this paper is to adopt BMF fuzzy neural network toadjust both the weightings of neural fuzzy networks and the controlpoints of BMFs. The traditional fuzzy neural network is trained byusing the learning algorithm with a gradient descent method, but thesolution of the fuzzy neural network maybe be fell into local minimumduring the learning process. While we build a random initial fuzzyevaluation rules, the sequential search-based single crossover pointmethod is adopted to implement the learning algorithm and the wholelearning structure can be established to adjust the weights and theBMF's. We use real genetic algorithm for searching the optionalsolution, and there are no encoding and decoding operations involvedduring the learning process. Finally, the simulation results show thatthe proposed approach has better results.