Based on analyzing fundamental algorithms principle of Back Propagation Network, in view of the limitations of BP network algorithm, this paper proposed the homologous improved algorithms from two respects: accelerate the learning speed of BP network and advance the convergence degree of network. By means of increasing the discrepancy factors of neuron function, the traditional function is optimized and improved, so that the output value break away from local minimal, greatly increasing the convergence speed of network; In the circle iterate expression, by means of increasing the momentum factor, it can filtrate the high frequency deviation of curved face in weight space, so as to widen the space of effective weight, and avoid the oscillation in the network learning process, enormously decrease the time of network learning.
展开▼
