臺灣博碩士論文加值系統

　　神經網路自從發展以來便廣泛的被大眾所使用，這也使得許多的學者投入研究各種神經網路演算算法的行列，使得各種的神網風路演算法如雨後春筍般的出現。而T.Kohonen在1973年提出了神經網路自組織圖（Self-Organizing Maps ）的演算法，在神經網路中的神經元之間加入了鄰接關係成為自組織圖演算法的特色，在神經網路演算法中開創了一個新的領域。
　　在經過不斷的研究與應用之後學者發現了Kohonen自組織圖的缺陷，就是固定的自組織圖鄰接關係。於是開始有人研究發展改良式的自組織圖演算法，企圖使自組織圖的鄰接關係突破固定的限制。而我們將在本論文中以圖學的角度去分析Kohonen 自組織圖的鄰接關係，並發展出兩套可以互相配合的動態式自組織圖演算法：『演化型自組織圖』與『交配型自組織圖』。
　　在演化型自組織圖演算法中，我們將神經鍵向量與鄰接關係的組合視為圖形，並發展出圖形的演化生長操作，應用在自組織圖中，使得演化型自組織圖可以動態的依據輸入資料的特性，而去改變自己的神經鍵向量數目與鄰接關係。所以我們可以不必再為初始化自組織圖鄰接關係而煩惱，因為演化型自組織圖演算法會依據狀況而調整自己。
　　而交配型自組織圖演算法則是提供另一種自組織圖演化的選擇，利用交配型自組織圖演算法，我們可以將兩個不同神經網路結構的自組織圖做基因演算法的改良型基因交配操作，使之產生結構與父代不同的子代自組織圖。利用交配型自組織圖，我們可以將演化型自組織圖已經固定在區域最佳解的結構重組，使其有機會離開區域最佳解。從生物學的觀點來看，演化型自組織圖演算法對比於交配型自組織圖演算法，就像生物的無性生殖對比於有性生殖。
　　由於演化型自組織圖演算法與交配型自組織圖演算法都具有自行動態改變結構的特性，所以除了可以使用在傳統自組織圖的應用領域中，也相當適合應用在任何具有動態環境特性的應用領域。例如目前正日漸發展的自主性代理人(automatic agent ），代理人可以在這兩種演算法的搭配使用下具有知識成長與演化修正的特性。而這便是目前傳統Kohonen自組織圖演算法所欠缺的。

　　Since the concept of neural network was introduced, there are many neural network models developed and used broadly in different research areas. In 1973, Kohonen proposes a new generation of neural network models, called Self-Organizing Map (SOM). The SOM algorithm adds one kinds of fixed neighborhood relations among neurons into regular neural nets and learns new patterns under such neighborhood constraints.
　　Although SOM has been proven to be effective in many applications, the fixedness of its neighborhood relations bring many inconveniences. This dissertation analyzes SOM from the viewpoint of graph theory and proposes a couple of dynamic SOM algorithms: “Evolutionary Self- Organizing Map algorithm ”and “Mate Self-Organizing Map algorithm”.
　　In “Evolutionary Self-Organizing Map (EXOM) algorithm”, the weight functions of neurons and the neighborhood relations are combined as a neighborhood graph. As the training data are provided, this neighborhood graph can be dynamically evolved accordingly. Thus the initial state of neurons is not so critical and the final structure of this neural net will be adjusted to fit the training data automatically.
　　Based on genetic algorithm, “Mate Self-Organizing Map (MSOM)”is proposed as another choice of SOM evolution strategies. MSOM mates two different neural nets according to their genes and breeds two children neural nets with different gene structures from their parents'''' . As genetic algorithm, such children neural nets of MSOM have chance to escape from local minimum (local-minimum -free property).
　　From biology viewpoint, ESOM is one kind of asexual reproduction as well as MSOM is a sexual reproduction. With these two evolution strategies we can make SOM grow as biological population. Such flexible neural net structure not only patches the weakness of traditional SOM algorithm, but also makes possible some applications with uncertainty environment. For example, both evolution algorithms can be used in agent models to let agents with different knowledge or properties evolve and mate to produce their nest generations in an multiple agent architecture.