Selective Data Collection Method for Deep Reinforcement Learning

Abstract

In deep reinforcement learning, reinforcement learning is responsible for interacting with the environment to produce data, and artificial neural networks are responsible for value function fitting. It is observed that artificial neural networks converged differently to different inputs, which, in our analysis, is due to imbalanced data. Therefore, we propose selective data collection to boost the quality of the data by then discarding the excess data. It has been proved experimentally that our method can significantly contribute to the convergence rate of the reinforcement learning algorithm.