A Study on How to Help Back-propagation Escape Local Minimum

One peculiarity impairing the performance of back-propagation is the presence of local minimum. In this paper, a local minimum detection algorithm called LMD is advocated. LMD bases on two kinds of constraint of local minimum to detect and escape local minimum. Before back-propagation trapping into local minimum, LMD will take measure to rectify the adjusting method of weights and biases. Combining LMD, back-propagation is improved to LMDBP algorithm in this paper. Results of experiment of back-propagation and LMDBP indicate that LMDBP can escape local minimum much more sufficiently than back-propagation algorithm.