This paper studies learning of the binary neural networks (BNN) and its relation the logical synthesis. The network has the sigum activation and can approximate a desired Boolean function if parameters are selected suitably. In a parameter subspace the network is equivalent to the disjoint canonical form of the Boolean functions. Outside of the subspace, the network can have simpler structure than the canonical form where the simplicity is measured by the number of hidden neurons. In order to realize effective parameter setting, we present a learning algorithm based on the genetic algorithm. Performing basic numerical experiments, the algorithm efficiency is confirmed. Also, this paper studies the dynamic binary neural networks (DBNN) and its learning of periodic binary sequence. The DBNN is constructed by applying the delayed feedback to a three-layer network. As an application of object, we study representative power electronics circuits. We propose the learning of teacher signal corresponding to switching signals of the inverter and matrix converter. We give a systematic analysis tool. The Gray-code-based return map that is useful to grasp basic characteristics of the DBNN such as the number of periodic orbits and their domain of attraction. We have confirmed storage of the teacher signal and automatic stabilization of the periodic orbits.

114 工学部 (旧) ・工学研究科 ＞ 学位論文 ＞ 修士論文