Research On Automatic Generation Of Neural Network Topology

Artificial Neural Network(ANN) technology has experienced repeated setbacks since the very beginning. It can be said that research on ANN is one of today's hot topics. The biological neuron system is the result of natural evolution; however, most ANN systems are result of human designing. Research on ANN topology automatic generation is not only for discovering new topology structures, but also contributing to the research of intellectual origins. Research of this thesis is mainly focused on two directions:one is how to simulate natural evolution process using evolution strategy to develop network structures; the second is how to guide the search direction by using search technology with heuristic information. we discovered both path as follows:1.Improved NEAT algorithm with pendulum balancing experiment. This algorithm can be applied for traditional ANN models.2.Binary Logic Neural Network(BLNN) was introduced with improvement for speeding up calculation, reduced weight and threshold searching space.3.Binary Logic Neural Network structure automatic generation algorithm was proposed:Binary Logic Neural Network structure generation based on Genetic Algorithm was implemented and tested with maze path covering experiment.4.A new way to generate BLNN by using heuristic information was proposed aimed at conquering the limitation of genetic algorithm. Proposed method can be used to find out the logic relationship between input and output and to generate the required structure for problem to be resolved.5.Finally, the application and limitation of each algorithm and future work was discussed.