Pulse waveform synthesis using recurrent complex-valued neural networks

An experiment of time-sequential pulse train synthesis using a layered and partially-recurrent complex-valued neural network is reported. A one half of the three-layer complex valued neural network is used to generate sinusoidal oscillation, and the other half is used to synthesize adaptively the intended pulse shapes and sequences. Stable time-sequential pulse signals are obtained after completion of the learning process. This result suggests that the complex-valued neural networks are effectively applicable to pulse-operational neural networks