長・中・短期記憶を用いた発見的最適化法

抄録

An optimization method is proposed for unconstrained optimization problems, including neural network training. To avoid falling into local optima, the method is based on stochastic search and performs diversified and intensified searches alternatingly. Besides, for more efficient search of the solution space, it utilizes its history of searching which is stored in long, medium and short term memories and controls the diversification and intensification of the search. The method includes minimum number of adjustable parameters aiming at a generic method that does not need much parameter tuning. An example shows that the method can find the global optimum irrespective to the initial solutions.