Deep Reinforcement Learning for Optimization at Early Stage

Abstract

We introduced Deep Reinforcement Learning(DRL) for design cost optimization at early stages of the System on Chips(SoCs) design process. We demonstrate that DRL is a suitable solution for the problem at hand. We benchmark three DRL algorithms based on Pointer Network, a neural network specifically geared for combinatorial problems, on the design cost optimization. We confirm the considerable improvements in cost optimization using DRL algorithms compared to conventional optimization methods. Additionally, by using the recently introduced RUDDER method and its reward re-distribution approach, we obtain a significant improvement in complex designs. Here, the obtained optimization is in average 15.18% on the area, 8.25% and 8.12% on the application size and execution time on industrial hardware/software interface designs

Publication
IEEE Design & TEST