A Data-Driven Automatic Design Method for Electric Machines Based on Reinforcement Learning and Evolutionary Optimization

Abstract

The design problems of electric machines are actually treated as a kind of mixed-integer problem, because the machine shapes are defined by some integer variables, such as number of slots, and the other variables, such as the tooth width, which are here called the fundamental and shape variables, respectively. To automatically solve these design problems, this article presents an automatic design method by combining the reinforcement learning and evolutionary optimization. In the proposed method, the decision process is modeled as a tree structure to seek for the fundamental variables, which are determined as a result of the tree search depending on the value functions of the nodes. Then, the shape variables are estimated from the function of the fundamental variables. These functions are constructed based on the design data, to generate which the reinforcement learning and evolutionary optimization are employed. As a result, the proposed method can automatically be adapted to unexperienced design problems through the data generation and function learning. The proposed method is applied to a design problem of a linear induction motor. It is shown that the machine designs with the prescribed performance for given specifications are automatically obtained. Moreover, it is also shown that the acceptable candidate designs can immediately be generated when the given specification is similar to the previously-solved problems by utilizing the design data generated by the past explorations.