Hardware-oriented deep reinforcement learning for edge computing

Abstract

A new deep reinforcement learning enhancement is proposed for edge computing. This work focuses on deep Q-networks (DQNs), which are used in deep reinforcement learning. Although DQNs are typically improved through a software-based approach, hardware-specific knowledge such as that on data paths and pipelines is used for improving a DQN. The DQN performance is improved and the number of resources are reduced through an efficient hardware design that considers the learning flow and parameter search. As the scale of the problem increases, the amount of reduction in the use of resources also increases. For example, when the size of the block catch game is 5 × 10, the memory requirement is reduced by approximately 50% compared to a previous DQN. The proposed hardware-oriented approach can be applied to any software technology. This study facilitates the development of novel technologies that can be realized through edge computing.