Fuzzy Neural Network For Construction Bidding Systems

The study of bidding decision-making is a very important subject in the present construction projects. Bidding decision-making needs to consider a large number of highly fuzzy, complex and interrelated factors. Due to the better self-learning and reasoning abilities of fuzzy neural networks based on T-S model (FNN-TS), it is used in the research of construction bidding.In FNN-TS, the consequent is a linear function of the input variables. And it allows expressing complicated behaviors with few rules. However, the structure without hidden layer in consequent subnetwork is too simple, which can't reflect the essence of systems. And the traditional method of generating fuzzy rules will cause the number of rules increasing in exponential way. Therefore, an improved fuzzy neural network based on T-S model (IFNN-TS) is proposed. In IFNN-TS, a hidden layer is added in consequent subnetwork of FNN-TS, and fuzzy rules are generated from training patterns then chosen depending on their importance. Whereas, the above procedure of choosing rules is arbitrary and subjective because of the membership function. So the subtractive clustering and the technique of determining the degree of applicability of each rule with Euclidean distance are used. And based on IFNN-TS, a modified neural network with subtractive clustering (NN-SC) is proposed. The fuzzifier and defuzzifier are needless in this model.The generalization ability of the networks relies on the pattern characters and the structure of the network. For enhancing the generalization ability, the training patterns are classified and processed using transcendent information of patterns and fuzzy inference, and the regularization is added which coefficient can be adjusted dynamically by fuzzy reasoning. To test the performance of the models and algorithm, they are used in construction bidding systems. Meanwhile, the improved BP neural network and the multi-input fuzzy neural network model are given to contrast with the models mentioned above. The simulation indicates that the methods in this paper are provided with rapid convergence, good learning and generalization abilities.