Evolving Neural Network Structure by Indirect Encoding Based on BQPSO
Fang Bao, Jun Jie Sun, Wenbo Xu · 2011
This paper proposes a novel algorithm of neural network structure evolve. First, the algorithm designs an indirect encoding schema representing the structure of neural network, use joint seed representing the existence of connection in neural network. Then, creating and evolving the coordinates of the joint seed using Binary Quantum-behaved Particle Swarm Optimization (BQPSO), evolving the value of the joint seed using nine-palace evolving rule, by separately evolve the coordinates and value of the joint seed, the growing and pruning of the network structure is achieved. The experimental results show that the algorithm has stable complexity when dealing with different scales of neural network. By the proposed indirect encoding schema and separated coordinates and value evolving, the algorithm solves the problem of geometrically growing structure-evolving complexity successfully.