自我組織特徵映射網路於最佳化問題及其應用

Abstract

自我組織特徵映射神經網路（SOM）已經廣泛地應用在靜態資料處理與動態資料的分析，但利用SOM解決最佳化的問題的研究非常少。目前以SOM為基礎的最佳化演算法對動態系統最佳化的效能還有待改進，所以在本論文中提出一自我組織特徵映射神經網路最佳化演算法（SOMS）應用於靜態與動態最佳化問題。為了更進一步擴展它的收尋能力，也提出一個新的鍵結值的調整規則以達到動態調整SOM的鄰域函數。在論文中我們也利用SOMS演算法發展ㄧ智慧型雷達預估器，可在很短的時間週期內對於只有很少的資料被雷達接收的情形下，估測目標物的運動軌跡。除此之外，當最佳化問題存在多個最佳解時，利用一個新的Niching方法（即決策型競爭機制），我們也提出一個Niching型自我組織特徵映射神經網路最佳化演算法（NSOMS）。為了提高學習的效能且同時可以讓最佳解的分佈結構顯現在二維的輸出空間，我們提出一新的神經元鍵結值與座標位置的調整規則，由於新的調整規則的設計簡單而且只用到兩個學習參數分別於神經元鍵結值與座標位置上，所以提出的NSOMS可以很容易地應用在不同的最佳化問題上。我們以模擬的方式來驗證此方法的可行性，並與傳統的卡曼濾波器 (KF)，基因演算法（GA），與其他SOM最佳化演算法進行比較。

The self-organizing map (SOM), as a kind of unsupervised neural network, has been used for both static data management and dynamic data analysis. To further exploit its search abilities, in this dissertation we propose an SOM-based search algorithm (SOMS) for optimization problems involving both static and dynamic functions. Furthermore, a new SOM weight updating rule is proposed to enhance the learning efficiency; this may dynamically adjust the neighborhood function for the SOM in learning system parameters. Based on the SOMS, we develop an intelligent radar predictor to achieve accurate trajectory estimation under the strict time constraint due to only few data are available in every short time period. Moreover, when an optimization problem has many different optimal solutions, a new niche method (deterministic competition mechanism) to extend SOM-based search algorithm (NSOMS) has been proposed for identification of multiple optimal solutions. The proposed NSOMS network structure is able to find multiple different optimal solutions and visualize distribution and structure of optimal solutions, allowing us to easily classify the optimal solutions into clusters. As a demonstration, the proposed NSOMS is applied for function optimization in a multimodal domain and also dynamic trajectory prediction involving multiple targets, with its performance compared with the genetic algorithm (GA) and other SOM-based optimization algorithms.