Estimation of Bayesian a posteriori probabilities with an autonomously learning neural network

We have previously devised an autonomously learning, adaptive neural network model for online classification and prediction tasks (C.P. Lim and R.F. Harrison, 1995). The system is based upon an integration of two neural network architectures: Fuzzy ARTMAP and the Probabilistic Neural Network. We demonstrate that the hybrid system is capable of providing outputs which estimate Bayesian a posteriori probabilities. It is also able asymptotically to achieve the Bayes optimal classification rates, online, without prior knowledge of impending changes in the environments. To assess the performance of the system, a benchmark problem of learning and predicting noisy patterns is presented, and the result is compared with other approaches.