FPGA-Based Acceleration of Reinforcement Learning Algorithm

Abstract

Reinforcement learning (RL) is a computational approach that trains agents in decision-making through iterative interaction with their environment. Despite RL's success in various sectors, like robotics, autonomous driving, and gaming, its intensive computational requirements limit its use in scenarios needing immediate response. Fieldprogrammable gate arrays (FPGAs), with their ability to perform diverse digital functions, come forth as a solution. FPGAs enhance RL by speeding up processes through parallel execution, reducing delays by integrating computations into hardware and offering reprogramming capabilities for diverse RL algorithms. This paper examines the current state of FPGAdriven advancements in RL, discusses the technological strides, challenges, and future potential, and highlights the use of FPGA accelerators in applying RL to real-time systems.