- 學位論文
混合型智慧系統之應用-使用邊界值解構樹法
Applications of Hybrid Intelligence System with Bound Decomposition Tree Algorithm
摘要
類神經網路擁有許多優點，譬如可以進行非線性模型的預測、可以經由訓練資料自行學習、高適應力等等…。所以已經被成功的應用在許多的領域，譬如機械控制、土木水利、氣象預測、財務分析、影音辨識等等…。但是由於類神經網路將所學習得到的知識儲存在它的神經鍵值跟網路架構中，其推導過程僅是一連串的數值運算，無法明確的解釋其所產生的推論結果，所以通常被稱為黑箱作業。然而對於預測的結果提供解釋的能力在某些領域是很重要的，如信用卡的審核跟醫療診斷等。相反地，專家系統將知識以規則的方式存在知識庫當中，並且利用邏輯推演來得到答案。其推演的過程比較接近人類思考方式，所以使用者可以檢驗專家系統的推演過程是不是正確。因此，也比較能夠信任專家系統的建議。可是專家系統的知識庫建立不容易，必須經由知識建構師(通常是電腦專家)與領域專家之間反覆多次的會晤溝通後，才能建立出適合的規則。而這個溝通與建立規則的過程通常要花上長久時間。而且許多領域專家經常無法明確的說明他們的領域知識，更造成知識擷取的困難。 因此，能夠從類神經網路找出對應規則的規則萃取演算法變得越來越重要。規則萃取演算法是一種方法，可以將類神經網路學習到的知識以規則的方式來呈現。有了這樣的演算法，我們可以建立一個混合式的智慧型系統。這個系統可以先利用神經網路模組經由訓練資料來學習並且得到知識。然後使用規則萃取演算法從學習好的類神經網路中將這些知識以規則的方式來呈現，並且利用這些規則來建立專家系統模組的知識庫。如此，這樣的混合式的智慧型系統可以同時擁有類神經網路與專家系統的優點。 在本篇論文中，我們設計了一個規則萃取演算法，能夠從類神經網路中找出規則並以布林邏輯或模糊規則的方式來表現。這個演算法是一種解構型的的規則萃取演算法，這類型的方法主要是針對各個神經元分別來取得規則。這個演算法主要的概念是經由快速的計算輸入集合的最大值與最小值來判別神經元是否會被激發。然後再反覆的將輸入集合分解成更小單元的集合，直到找出集合中的輸入全部都會使神經元激發或者全部都不會為止。由於整個解構的過程如同二元樹一樣，所以這個演算法命名為「邊界值解構樹法」(Bound Decomposition Tree Algorithm, BDT)。 這個演算法得原理非常簡單但是卻很有效率。同時，演算法中擁有一個規則簡化的方法能夠同時增進效率與降低規則的複雜度。在論文中，用了一個線性方程式與三個機械學習領域常用的資料庫(MONK’s、大選投票資料跟蘑菇分類)來測試這個演算法，並且跟MofN與Garcez兩種來進行比較。藉由這個演算法，我們也建立了一個混合式的智慧型系統。並且在過去幾年中，嘗試著將這個混合型的智慧型系統使用在幾個實際的應用案例中（如網頁計數管理、中西醫冠心病輔助診療系統、人才適任評估系統與PECVD的機台參數預測與調整建議系統），並且獲得一些初步的成功運用。這些案例也會在本論文中進行探討。
並列摘要
Neural network has many advantages such as non-linear module prediction, automatic learning from training data, high adaptation et al. It also is successfully applied to many applications in different domains, like control system, civil engineering, weather prediction, financial analysis and recognition of image and speech et al. However, due to the results made by the neural networks are difficult to explain the decision process of neural networks is supposed as a black box. This is because the learned knowledge is stored in the weights and structure of the network and the decision process of neural networks is a numerical value operation. But the explanation of decision making is important to some applications such like credit approval application and medical diagnosing software. Oppositely, expert system stores the knowledge in the knowledge base with symbolic logic rule form and inferences with this logic rules. Hence, the decision making process of expert system is close to human thought. By this, users also can verify the conclusion made by the expert system, and then they are more likely to trust the results. But it is not easy to build the knowledge base of the expert system. The acquiring knowledge process is needed a frequently and lengthily interview between a knowledge engineer and domain experts, and the process may takes a lone time. Moreover, many domain experts could not explain their domain knowledge explicitly. Therefore, rule extraction algorithm is becoming more and more important in explaining the knowledge from the neural networks. Rule extraction algorithm is a method to get rules from a neural network and these rules can show the learned knowledge within the neural network. With this kind of algorithm, we can build a hybrid intelligence system. The system can use neural network module to learn the knowledge from training data automatically and then transforms the knowledge to symbolic rules. Next, it uses these rules to construct the knowledge base of expert system module. Therefore, the hybrid intelligence system can have the advantages of both neural network and expert system. In this thesis, a rule extraction algorithm is analyzed and designed to extract rules from neural networks. The algorithm is a kind of decompositional rule extraction algorithm, and it means the algorithm extract rules from each neuron independently. The notion of the algorithm is designing a quick equation to find the maximum and minimum values of a set of inputs of a neuron, and then it can observe if all of these inputs in the set will make the neuron activated (or all not). Next, the input set will be decomposed into smaller subsets until all the inputs in these subsets will make the neuron activated (or all not). Because the decomposition process is like a tree, the algorithm is named as “Bound Decomposition Tree Algorithm” (BDT). The algorithm is simple and efficient and it also can reduce the extracted rules. The reduce option of the algorithm can make simple extracted rules, reduce the number of these rules and improve the efficiency. In the thesis, we use a linear equation and three familiar data used in machine learning (MONK’s, the voting data, and mushroom data) to test the algorithm and compare with other two algorithm, MofN and Garcez. Using the algorithm, we also build a hybrid intelligence system which combines advantages of neural network and expert system. In the last years, the hybrid intelligence system was applied into some applications, such as “abnormal web access”, “the expert system of coronary artery disease in Chinese and western medicine”, “employee recruitment with Biodata” and “advance process control for PECVD”. These applications will also been discussed in the thesis.