Study On Neural Network Ensemble And Its Application In Soil Science

Artificial neural network (ANN) has been applied in many fields and achieved plentiful fruits already. However, a practical problem that has come to prominence is that ANNs are unstable predictors. Neural network ensemble is a learning paradigm where a collection of a finite number of ANNs is trained for the same task. It originates from research work of Hansen and Salamon, which shows that the generalization ability of a neural network ensemble system can be significantly improved through ensembling a number of ANNs. Since it behaves remarkable well, it has become a very hot topic in both neural networks and machine learning recently.In this dissertation, the optimization for combining of component ANNs' results in an ensemble is explored with genetic algorithm and dynamic aspect in regression problems, and this technology is also applied to the spatial distribution of soil properties, which is very popular in soil science recently. The main contributions of this dissertation are summarized as follows:Firstly, it is effective to optimizing the integrated weights of component ANNs with genetic algorithm by investigation of weighted optimization with genetic algorithm thoroughly. The approach that choose component ANNs by genetic algorithm directly, instead of optimizing integrating weights and then selecting component ANNs, is also effective. This approach is useful and simple for selective ensemble with genetic algorithm.Secondly, a dynamic weighted ensemble approach is proposed. The dynamic weighted integration is realized by generalized regression neural network (GRNN). Experimental results show that it outperforms the traditional ensemble approaches.Thirdly, a new approach which combines the selective method and dynamic idea is proposed. It includes two steps: selection of component ANNs by genetic algorithm at first and dynamic integration by GRNN at last. It is effective proved by the experiments.Fourthly, the neural network ensemble technique is applied in the spatial distribution of soil properties. As compared to kriging method, the neural network ensemble achieves better or similar accuracy of prediction and estimated contour maps. The potential ability of neural network ensemble for soil spatial variety is good.Fifthly, it is benefit to optimize the integrated weights of component ANNs in traditional Bagging and Boosting.