Constructive deep deterministic policy-based reinforcement learning method for optimization of DDR5 memory signaling architecture

Abstract

In this article, we propose for the first time a constructive deep deterministic policy-based reinforcement learning method for optimizing DDR5 memory signaling architectures. The proposed method is a design optimization methodology for maximizing the eye diagram, which is a performance indicator of the channel for many factors such as desired components and termination and decision feedback equalizer, in the entire channel from the processor to the memory in order to improve the limitations of the DDR5 channel. Deterministic policy is used to optimize successive factors, and deterministic policy-based reinforcement learning based on constructive neural networks is used to reflect the association of many factors being optimized. The proposed method can be reused because the result can be obtained immediately even if the values of a given environment are changed. To verify the optimality and optimization time of the proposed methodology, we compared the lattice search method based on the Latin hyperbolic sampling method, the random search method, and the conventional optimization methods such as the Bayesian optimization method, and probabilistic policies such as the advanced actor critical and proximity policy optimization. We compared it with the reinforcement learning method based on it. As a result, the proposed method proved to be the best method in terms of optimization performance, optimization time, and reusability.