Osnabrueck University | overview of specific architectures in the field of reinforcement learning

【 title 】A Survey of Domain-Specific Architectures for Reinforcement Learning

【 The author team 】Marc Rothmann, Mario Porrmann

【 Date of publication 】2022.1.26

【 Thesis link 】https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9694573

【 Recommended reasons 】 Reinforcement learning algorithm has achieved great success in solving continuous decision-making problems in many different problem areas . However , Their training is usually time-consuming , The training time ranges from a few hours to a few weeks . The development of reinforcement learning domain specific architecture is expected to speed up computing , Reduce experimental turnaround time , Improve energy efficiency . This paper summarizes the hardware structure of reinforcement learning algorithm acceleration . be based on FPGA The realization of is the focus of this work , But also based on GPU Methods . This review includes tables and deep reinforcement learning algorithms . The technologies used in different implementations are introduced and compared . Last , Based on the previous discussion of the existing architecture , The possible areas of future work are put forward .