行政院國家科學委員會專題研究計畫 成果報告
演化式模糊類神經推論系統於營建管理之應用(I)
計畫類別： 個別型計畫
計畫編號： NSC91-2211-E-011-060-
執行期間： 91 年 08 月 01 日至 92 年 07 月 31 日 執行單位： 國立臺灣科技大學營建工程系
計畫主持人： 鄭明淵
計畫參與人員： 陳弼宏、林家豪
報告類型： 精簡報告
處理方式： 本計畫涉及專利或其他智慧財產權，1 年後可公開查詢
中 華 民 國 92 年 12 月 25 日
行政院國家科學委員會專題研究計畫成果報告
演化式模糊類神經推論系統於營建管理之應用(I) Evolutionary Fuzzy Neural Inference System (EFNIS) for
Construction Management (I) 計畫編號：NSC 91 –2211-E-011-060 執行期限：91 年 8 月 1 日至 92 年 7 月 31 日
主持人：國立台灣科技大學營建系 鄭明淵教授 共同主持人：
計畫參與人員：國立台灣科技大學營建系 陳弼宏、林家豪
中文摘要
營建管理領域的許多問題具有複雜、
不確定以及隨環境變動的特點，而基因演 算法、模糊理論與類神經網路在許多文獻 中也已分別成功的應用在營建管理領域中 解決各種不同的問題。此三種方法具有優 缺點互補的特性，因此，本研究考量上述 三種人工智慧技術的優點與特性，將基因 演算法、模糊理論與類神經網路三者結 合，進而發展一「演化式模糊類神經推論 模式」來解決營建管理中隨環境變動的複 雜以及不確定的問題。「演化式模糊類神 經推論模式」係模擬人類決策模式與行 為，所發展之智慧型學習推論機制，模式 中應用基因演算法搜尋所有模糊類神經網 路所需要的最佳參數，例如：模糊理論中 隸屬函數的訂定及搜尋最佳類神經網路架 構；透過模糊理論處理近似推論與不確定 性問題；而類神經網路則用於學習尋找輸 入變數與輸出變數間的關係，因此，本模 式可藉由工程經驗學習累積，自動求得適 應環境的最佳解。
關鍵詞：基因演算法、模糊理論、類神經 網路、演化式模糊類神經推論模 式
Abstract
Most problems in construction management are complex, full of uncertainty, and vary with environment. Genetic Algorithms (GAs), Fuzzy Logic (FL), and Neural Networks (NNs) have been
successfully applied in construction management to solve various kinds of problems. These three computing methods offset the demerits of one paradigm by the merits of another. Considering the characteristics and merits of each method, this research combines the above three techniques to develop an Evolutionary Fuzzy Neural Inference Model (EFNIM). In the model, GAs is primarily concerned with optimization; FL with imprecision and approximate reasoning; and NNs with learning and curve fitting; Thus, the best adaptation mode is automatically identified.
Keywords: Genetic Algorithms (GAs), Fuzzy Logic (FL), Neural Networks (NNs), Evolutionary Fuzzy Neural Inference Model (EFNIM)
RESEARCH MOTIVATION
Construction engineering is made up of a set of different activities. Those activities interact with each other and are affected by a variety of uncertainties such as weather conditions, geological properties, human judgments, random market fluctuations, etc.
Appropriate construction management is a necessary process for construction engineering to efficiently accomplish construction objectives. The success of projects depends on construction management [1].
Construction management is a process of accomplishing construction objectives by
construction resources. Due to the uncertainty and changing nature of construction industry, most practical construction management problems are complex, uncertain, and vary with the environment [2]. Developing a deterministic mathematical model for solving such construction management problems is difficult and expensive. Thus, approximate inference, which is fast and cost effective is an alternative to solve the problems in construction management.
Inference is the process of deriving new facts from previously known facts. When the previously known facts change, the inference process should be adapted accordingly. Construction management problems that vary with the surroundings are complex with full of uncertainties, vagueness and have inexact and incomplete information, the previously known facts change continuously. Hence, the inference process should be adapted with the real world environment [3]. Human is able to learn and is capable of processing complex problems with uncertainty, imprecision, and incomplete information in the changing vicinity. Imitating the process of human inference is effective for solving the construction management problems.
Artificial intelligence (AI) is concerned with building computer systems that solve the problem intelligently by emulating the human brain. AI technology provides techniques for the computer programs to carry out a variety of tasks, at which humans are currently better [4]. Consequently, AI paradigms are appropriate for solving the complex and uncertain construction management problems that vary with the changing environment [5].
The most popular AI paradigms are fuzzy logic (FL), neural networks (NNs), and genetic algorithms (GAs). The combination of FL, NNs, and GAs offset the demerits of one paradigm by the merits of another [6].
In the last few years, several articles have been devoted to the study of fusing FL, NNs, and GAs to derive a better model performance than those using a single conventional method [7]. In general, in real
world applications, FL is used to define system inputs/outputs by fuzzy sets or to infer with inference principle [8]; NNs to tune the shapes of membership functions (MFs) or to extract fuzzy rules from training patterns [9]; and GAs to search for the optimum interconnections, weights between layers, number of hidden neurons of NNs, or distribution of MFs of FL [10]. However, most techniques treat these parts separately, generate some illegal solutions, or restrict NNs with fixed layers for different problems resulting in sub-optimal solutions [11-13].
RESEARCH OBJECTIVES
The primary objective of this research work is to develop an intelligent inference system for solving the complex and uncertain construction management problems that vary with the environment. In order to achieve the primary goals, this research fuses FL, NNs, and GAs to develop an Evolutionary Fuzzy Neural Inference Model (EFNIM) that simultaneously searches for the fittest distribution of MFs with optimum and legal network topology with optimum network parameters.
MODEL ARCHITECTURE
The architecture of the EFNIM is shown in Fig. 1. The proposed EFNIM is a fusion of FL, NN, and GA paradigms. The combination of FL, NNs, and GAs offset the demerits of one paradigm by the merits of another. In the formulated model, FL is primarily concerned with imprecision and approximate reasoning; NN with learning and curve fitting; and GAs with optimization.
Input patterns Knowledge
Fuzzifier Learning engine Defuzzifier (NN)
MF
NN’s parameter and topology Optimization
(GA)
FIG. 1. EFNIM Architecture
The objective of this research is to
develop an intelligent inference model.
Therefore, the model is developed based on FL that mimics the high level of human inference process. FL and NNs are complementary technologies. The combination of these two technologies into an integrated system appears a promising path towards the development of intelligent systems capable of capturing qualities characterizing the human brain. In Fig. 1, fuzzy inference engine and fuzzy rule base in the traditional fuzzy logic system are replaced by the NN. The NN is used to overcome the difficulties in acquisition of fuzzy rules and determination of composition operator and to offer a learning ability to the integrated system. The combination of the FL and NN is regarded as a “neuro with fuzzy input-output,” which is also a neural network with both fuzzy inputs and fuzzy outputs. In this work, for convenience to describe the “neuro with fuzzy input-output,”
it is initialized by the FNN which is a general phrase to express fusion/union of FL and NN [4].
Although the FNN is more reasonable than traditional FL to simulate the characteristics and process of human inference, the FNN for learning different tasks has demonstrated the difficulty in selecting an appropriate topology as well as appropriate parameters for a network. In addition, the determination of suitable distribution for the MFs, for solving disparate problems is time consuming and the difficulty increases with problem complexity.
GA is an effective approach to conquer the drawbacks of FNN [8]. Therefore, the EFNIM employs GA to simultaneously search for the fittest shapes of MFs, optimum FNN topology, and optimum parameters of FNN.
MODEL ADAPTATION PROCESS
The adaptation process of the EFNIM is described by the following pseudo code algorithm. In the algorithm, P(t) denotes a population of ξ individuals at generation t, PO(t) denotes an offspring population of σ individuals, and PM(t) denotes a mutation
population of τ individuals.
Begin t ← 0;
Initialize population P(t);
Evaluate every individual in P(t);
Do while (not termination condition) Perform crossover in P(t) to yield PO(t);
Perform mutation in P(t) and PO(t) to yield PM(t);
Evaluate every individual in PO(t) and PM(t);
Select P(t＋1) from P(t), PO(t), and PM(t);
t ← t＋1;
Loop End
CONCLUSIONS
This research fuses GAs, FL, and NNs to develop an EFNIM. In the model, FL is primarily concerned with imprecision and approximate reasoning; NNs with learning and curve fitting; and GAs with optimization.
The proposed model searches all possible shapes of MFs, NN topologies, and parameters of NNs including summit positions of MFs, widths of MFs, number of hidden layers, number of neurons per hidden layer, interconnections, synaptic weights, bias values, and slopes of activation functions. The characteristic of the model is to simultaneously achieve the fittest distributions of MFs with minimum network topology and optimum parameters of FNN.
The present research work pioneers the applications of fusing FL, NNs, and GAs in construction management. It promotes the AI technology in construction industry.
REFERENCES
[1] Bush, V. G. (1973). Construction management: a handbook for contractors, architects, and students. Reston, Reston, 1－6.
[2] Li, H. (1996). “Case-based reasoning for intelligent support of construction negotiation.” Information
& Management, 30(5), 231－238.
[3] Mareels, I., and Polderman, J. W. (1996). Adaptive systems: an introduction. Birkhauser, Boston, 1－
3.
[4] Haykin, S. (1999). Neural networks: a comprehensive foundation. Prentice-Hall, New
York.
[5] Tommelein, I. D., Levitt, R. E., and Hayes-Roth, B.
(1992). “Site-layout modeling: how can artificial intelligence help?.” Journal of Construction Engineering and Management, ASCE, 118(3), 594－611.
[6] Martin, N. M., and Jain, L. C. (1999).
“Introduction to neural networks, fuzzy systems, genetic algorithms, and their fusion.” In Jain, L. C.
and Martain, N. M. eds., Fusion of neural networks, fuzzy sets, and genetic algorithms:
industrial applications, CRC Press, Boca Raton, 3
－12.
[7] Linkens, D. A., and Nyongesa, H. O. (1996).
“Learning systems in intelligent control: an appraisal of fuzzy, neural and genetic algorithm control applications.” IEE Proceedings: Control Theory and Applications, 143(4), 367－386.
[8] Gorzalczany, M. B., and Gradzki, P. (2000). “A neuro-fuzzy-genetic classifier for technical applications.” Proceedings of IEEE International Conference on Industrial Technology, 1, 503－
508.
[9] Ghezelayagh, H., and Lee, K. Y. (1999). “Training neuro-fuzzy boiler identifier with genetic algorithm and error back-propagation.” IEEE Power Engineering Society, Summer Meeting, 2, 978－982.
[10] Jagielska, I., Matthews, C., and Whitfort, T.
(1999). “An investigation into the application of neural networks, fuzzy logic, genetic algorithms, and rough sets to automated knowledge acquisition for classification problems.”
Neurocomputing, 24, 37－54.
[11] Liska, J., and Melsheimer, S. S. (1994).
“Complete design of fuzzy logic systems using genetic algorithms.” Proceedings of the Third IEEE Conference on fuzzy systems, 2, 1377－
1382.
[12] Ishigami, H., Fukuda, T., Shibata, T., and Arai, F.
(1995). “Structure optimization of fuzzy neural network by genetic algorithm.” Fuzzy Sets and Systems, 71(3), 257－264.
[13] Qin, K., Wang, W., and Gong, M. (1997). “A genetic algorithm for the minimum weight triangulation.” IEEE International Conference on Evolutionary Computation, 541－546.
行政院國家科學委員會補助專題研究計畫成果報告
※※※※※※※※※※※※※※※※※※※※※※※※※※
※ ※
※ 演化式模糊類神經推論系統於營建管理之應用(Ι) ※
※ ※
※※※※※※※※※※※※※※※※※※※※※※※※※※
計畫類別：█個別型計畫 □整合型計畫 計畫編號：NSC－91－2211－E－011－060 執行期間：91 年 8 月 1 日至 92 年 7 月 31 日
計畫主持人：鄭明淵 教授 共同主持人：
計畫參與人員：陳弼宏、林家豪
本成果報告包括以下應繳交之附件：
□赴國外出差或研習心得報告一份
□赴大陸地區出差或研習心得報告一份
□出席國際學術會議心得報告及發表之論文各一份
□國際合作研究計畫國外研究報告書一份
執行單位：國立台灣科技大學營建系
中 華 民 國 92 年 7 月 31 日
