Using Training Curriculum with Deep Reinforcement Learning. On the Importance of Starting Small

Abstract

Reinforcement learning algorithms are being used to solve problems with ever-increasing level of complexity. As a consequence, training process becomes harder and more computationally demanding. Using transfer learning can partially elevate this issue by taking advantage of previ-ously acquired knowledge. In this paper we propose a novel test environment and experimentally evaluate impact of using curriculum with deep Q-learning algorithm.