Research on the learning algorithm of BP neural networks embedded in evolution strategies

Combining the BP algorithm and evolution algorithm, the gradient-BP algorithm (EBP) was proposed. Because BP algorithm used the idea of gradient descent, it was unavoidable that local minimalism unperfected exists. Evolution algorithm was a technique that simulates the evolution of animal. We introduced evolution algorithm into BP algorithm and it formed an evolution-BP algorithm. EBP algorithm absorbed nonlinear information of the error function, and it didn´t depend on the gradient information of target function. It not just improves the speed of local constringency, but also has the ability of global constringency. It avoids the possibility of local minimalism, and improves model´s precision and the speed of calculation.