A Reduction Method For Artificial Neural Network Inputs Based On An Improved Genetic Algorithm

Artificial neural network model is one of the most widely used classifier in the pattern recognition area. In order to obtain a high classifier, the feature reduction should be carried on to the training data used to build a neural network model, so that the redundant features and the irrelevant features are removed. The reduced training data can not only optimize the neural network's structure, but also reduce the jumbled hidden level nodes. The greatly most important is that it can significantly increase neural network's training speed and decrease neural network's classification error rate.The best feature set problem is a NP-hard problem, which may cause combinatorial explosion with the increase of the number of features, and result in exponential increasing in time complexity. As a result, the traditional optimized method is not applicable. To deal with this problem, a general genetic algorithm based feature combination optimization method is designed. This method regards the feature combination as an individual, and forms a population with certain individuals; by providing operations such as selection, crossover, and mutation, it obtains the superior feature combination through evolution from parent generation to child generation.To overcome the initial population is far away from the superior individual, and the convergence rate becomes slower, the feature combination with bigger information gain is assigned with a larger choice probability when the population is initialized according to the calculated information gain of each feature, considering that information gain reflects its discrimination ability with regard to the categorical feature of the data set. To save the fitness value computation time, the Hash table is utilized to store the individuals with different genetic series with each other, so that the evaluation speed can be enhanced further.14 machine learning UCI public data sets are used to carry out test experiments. Each data set is tested for 10 times and each time 10-fold cross validation is accomplished. The final error ratio is the mean error ratio of the 10-time experiment results.The experiments indicate that under the information gain inspiration, the initialization population can effectively improve the performance of the genetic algorithm, enhance the overall fitness level of the initial population, speed up the convergence rate of the genetic algorithm, obtain the superior feature combination more fast, and thus optimize the neural network's structure, raise its training speed, and reduce its classification error.