Decision boundary learning based on particle swarm optimization

In neural network (NN) learning, we usually find an NN to minimize the approximation error for a given training set. Depends on the data given, the performance of the NN can vary significantly. In fact, if the training data are close to the true decision boundary (DB), the NN can generalize well. On the other hand, if the given data are far away from the true DB, the DB formed by the NN can be very different from the original one, and the genelization ability of the NN cannot be high. Based on this observation, we propose a new concept called decision boundary learning (DBL) in this study. A direct way for DBL is to approximate the true DB using a support vector machine (SVM), and then find a set of training patterns using particle swarm optimization (PSO). Experimental results on four public databases show that the training patterns so obtained may generate much better NNs, and in all cases, the NNs are comparable to or better than SVMs, although they are much simpler.