A Neuro-Symbolic Approach to Runtime Optimization in Resource Constrained Heterogeneous Systems

Abstract

Our approach to runtime optimization of heterogeneous systems using Reinforcement Learning (RL) and neuro-symbolic Logical Neural Networks (LNN) is described. We provide details of a method for optimizing the runtime allocation of power and processor resources in heterogeneous System-on-Chip (HSoC) applications to maximize their performance/watt. We demonstrate promising results in creating human-interpretable and -interactable AI policy, learning with less data by incorporating domain knowledge, and use of logical reasoning to extend the optimization window beyond the training set. Register Transfer Level (RTL) HSoC circuit simulation results are used to show the comparative advantage of the neuro-symbolic policy implementation versus hardware power management.