방향 벡터와 스텝 사이즈 학습을 통한 켤레 기울기 최적화 알고리즘

초록

Gradient descent method is the most widely used first-order iterative algorithm for solving optimization problems. The first-order algorithm is based on the steepest descent method, and the algorithms to increase the rate of convergence through as conjugate gradient method has been studied. These methods use the information of the current state and the last iteration to determine the search direction. However, these methods do not use make good use of past information from the process of iteratively searching optimal point. Among the heuristic algorithms, reinforcement learning based zeroth-order algorithm is developed that makes an approximation model using historical design change actions and predicts the next action. In this paper, we propose a first-order optimization algorithm based on the learning of historical data. The proposed method uses the design change information from the past iterations to update the direction vector with acceleration term. Also, learning rate and acceleration parameter are learned based on historical data. The mathematical examples are performed to compare with existing methods and verify the performance of the proposed method.

키워드