Evolving Chaotic Neural Network for Creative Sequence Generation

This paper describes an approach to generate a sequence requiring an unrealizable function by programs, such as a flash that is required especially in creative activity of a human. We have already proposed a recurrent neural network that demonstrates a generation of several creative sequences, but convergency and stability problems occur. On the other hand, it is known in biological experiments where the chaotic sequences can be observed from brain waves. The neural network constructed from chaotic neurons has nonlinear dynamics, but there remains the difficulty of training method. We propose an evolutional methodology to train a chaotic neural network, and introduce Darwinism for its evolving process. To determine their most suitable structure and the weights of connection, we use AIC for the fitness value.