A Data-Driven Optimization Method for Strongly Non-Separable Mixed-Integer Problems

This paper presents a data-driven optimization method based on tree search-based reinforcement learning to solve strongly non-separable mixed-integer problems. With this method, some variables that mainly affect performance and non-separability are stochastically determined through a modified tree search. Then, other variables are stochastically estimated and optimized by evolutionary algorithms. These probabilities are adapted with a reinforcement learning approach. As a result, the present method makes it possible to easily solve non-separable mixed-integer problems. Moreover, in the reinforcement learning, previous optimization results can be utilized in the next optimizations. The present method is applied to a design problem of an induction motor, which is one of the practical engineering and non-separable mixed-integer optimization problems. It is shown that the present method can solve the design problem under various conditions.