<ul><li>Simulation-based design space exploration (DSE) aims to optimize high-dimensional structured designs efficiently under complex constraints and expensive evaluation costs.</li><li>CORE is a constraint-aware, one-step reinforcement learning method introduced for simulation-guided DSE.</li><li>CORE utilizes a policy agent that learns to sample design configurations, incorporates dependencies via a scaling-graph-based decoder, and penalizes invalid designs.</li><li>CORE focuses on hardware-mapping co-design of neural network accelerators, demonstrating improved sample efficiency and better accelerator configurations compared to existing methods.</li></ul>

CORE: Constraint-Aware One-Step Reinforcement Learning for Simulation-Guided Neural Network Accelerator Design

Discover more