Proposing ε-greedy Reinforcement Learning Technique to Self-Optimize Memory Controllers
Aritra Kumar Ray, Hena Ray · 2021 2nd International Conference on Secure Cyber Computing and Communications (ICSCCC) · 2021
This paper proposes the novel algorithmic idea of ε-greedy based Reinforcement Learning (RL) technique to self-optimize memory controllers. It aims to overcome the limitations of the existing RL based self-optimization of memory controllers by significantly reducing the space complexity and power requirements, while considering all the formulation constraints of the aimed problem. The proposed RL based DRAM scheduler is in line with the asymptotic correctness solution approach and does long-term planning so as to maximize the utility of the bus in the long run, while balancing exploration versus exploitation, providing room for interleaving of requests from different cores, optimizing the read and write requests and allowing precedence considering the criticality of each request. We have mathematically explained in this paper, and thereafter formulated the detailed algorithm, to exhibit how our proposed design of the RL based DRAM scheduler would bring about significant improvement in the existing design.