Construction of PSE System for Developing Reinforcement Learning Algorithms

Abstract

Many researchers who study stochastic simulations such as reinforcement learning algorithms have to run their programs many times to compare developing algorithms and find better sets of parameters for their programs. In order to reduce their working time, we build a problem solving environments (PSE) system to assist them. Our system has three subsystems: a distributed computing system, a data management system and a graph generation system. In this paper, we present a relationship between developing algorithms and the three subsystems. Using our system, users register their programs, run them on a distributed computing system, obtain results automatically, and compare them graphically. We conduct experiments with human subjects. As a result, a user obtained five times speedup for his work time through executions, screening simulation data sets and comparing algorithms. As these subsystems for PSE system, it is important not only distributed computing but also supporting both data management and graph generation.