QTAccel: A Generic FPGA based Design for Q-Table based Reinforcement Learning Accelerators

摘要

Q-Table based Reinforcement Learning (QRL) is a class of widely used algorithms in AI that work by successively improving the estimates of Q-values - quality of state-action pairs, stored in a table. They significantly outperform Neural Network based techniques when the state space is tractable. Fast learning for AI applications in several domains (such as robotics), with tractable `mid-sized' Q-tables, still necessitates performing a large number of rapid updates. State-of-the-art FPGA implementations of QRL do not scale well with the increasing Q-Table state space. Thus, they are not efficient for such applications. In this work, we develop a novel FPGA based design of QRL and SARSA (State Action Reward State Action), scalable to large state spaces and thereby facilitating a large class of AI applications. Our architecture provides higher throughput while using significantly fewer on-chip resources. It is capable of supporting a variety of action selection policies that covers Q-Learning and variations of bandit algorithms and can be easily extended for multi-agent Q learning. Our pipelined implementation fully handles the dependencies between consecutive updates allowing it to process one sample every clock cycle. We evaluate our architecture for Q-Learning and SARSA algorithms and show that our designs achieve a high throughput of up to 180 million samples per second.