Exploring the parameter space of a genetic algorithm for training an analog neural network
Steffen Hohmann, Johannes Schemmel, Felix Schürmann, Karlheinz Meier · 2002
This pap er presents exp erimental results ob-tained during the training of an analog hard-ware neural network. A simple genetic al-gorithm is used to optimize the synaptic weights. The parameter space of this algo-rithm has b een intensively scanned for two learning tasks (4 and 5 bit parity). The re-sults provide a quantitative insight into the interdep endencies of the evolution parame-ters and how the optimal settings are pre-determined by the learning problem. It is observed that p opulation sizes in the order of 15 in connection with mutation rates of ab out 1 % yield the b est p erformance of the training