実験に基づく多目的最適化における失敗試行を考慮した最適化アルゴリズムの開発

抄録

Simultaneous optimization with respect to multiple conflicting criteria is required in many fields of engineering including robotics. Such optimization in robotics often has to be achieved through experiments, which are expensive in time and/or money. Therefore, the evaluations of objective functions are scarce resources. In some cases, there are input regions on which the objectives cannot be defined, and which are unknown in advance. The existence of such unknown failure regions is a major problem in making an efficient experimental plan. This paper proposes a multiobjective optimization method which is capable of utilizing the unsuccessful samples appropriately to avoid further unnecessary experiments. By using Gaussian process classifier, the proposed method estimates the probability to fail at any inputs and use it to reduce the number of the unsuccessful evaluations. The efficacy is shown by numerical tests and a robotic experiment.