Bayesian Training of Artificial Neural Networks Using MCMC-PSO with Applications in Time Series

Training an ANN from Bayesian viewpoint is doneusing optimization of a cost function with PSO. Predicting newtime steps needs to accurately approximate posterior distributionof the data conditioned by network parameters. It has been doneusing a Markov Chain with this posterior distribution at thisequilibrium state and the state should be reached with MonteCarlo algorithm. Bayesian learning provides a natural method tohandle the model complexity. The proposed approach canestimate the efficient models that results a reasonablegeneralization because overfitting to training data is penalizedautomatically. Hence, the estimated models produce much betterforecasting performances. Moreover, derivative information isnot required comparing to gradient based approaches.

کلیدواژه ها:

Monte Carlo ، Markov Chain ، Neural Network ، Bayesian Learning ، PSO