AutoRL-TSP-RSM: automated reinforcement learning system with response surface methodology for the traveling salesman problem

Abstract

The tuning of parameters is an important step towards the use of machine learning methods. However, it can be costly to define these initial condition values for each application. Thus, this paper aims to propose an Automated Machine Learning system for parameter tuning. In this line, an Automated Reinforcement Learning method was developed applied to the Traveling Salesman Problem. The proposed system adjusted through the Response Surface Methodology two parameters (learning rate and discount factor) of the Q-learning algorithm. The results revealed that the values adjusted by the proposed method reached, in general, the best solutions, in comparison with the adoption of parameters from the literature.