Reproduce “Deep Reinforcement Learning for System-on-Chip: Myths and Realities” on OpenResearch

About this paper

Tegg Taekyong Sung, Bo Ryu

Researchers at EpiSys Science, Inc. developed a deep reinforcement learning (DRL) framework for System-on-Chip (SoC) resource allocation, introducing the Eclectic Interaction Matching (EIM) technique to address challenges like variable action sets and delayed rewards. Their work established that while DRL with EIM excels in balanced communication-computation scenarios, traditional heuristics often perform better in highly computation-intensive environments.

Key insights

• Direct transfer of existing neural schedulers from cloud cluster computing to the SoC domain is largely ineffective due to fundamental differences in resource heterogeneity and the dynamic nature of task action sets.
• The Eclectic Interaction Matching (EIM) technique is crucial for enabling DRL agents to learn in complex SoC environments, effectively addressing the challenges of variable action sets and delayed reward attribution.
• The Communication-to-Computation Ratio (CCR) is a pivotal operational condition, significantly influencing whether neural or heuristic schedulers achieve superior performance in SoC resource allocation.

Summary by alphaXiv.