Reproduce “Reinforcement Learning for Scalable Logic Optimization with Graph Neural Networks” on OpenResearch

About this paper

Xavier Timoneda, Lukas Cavigelli

Researchers at Huawei Technologies, Zurich Research Center, developed a scalable deep reinforcement learning framework leveraging Graph Neural Networks for logic optimization on Majority-Inverter Graphs. This approach introduces a novel graph-size-independent node embedding and a multi-action scheme, achieving up to 1.75 times greater size reduction compared to the ABC tool on 50-node random logic graphs.

Key insights

• Scalable node embeddings that are independent of global graph size are essential for applying GCN-based RL effectively to large-scale logic optimization problems.
• Enabling the simultaneous application of multiple local actions within an RL framework significantly improves the computational efficiency for graph transformation tasks.
• Integrating automatic, always-applied inference rules into the environment can simplify the RL agent's search space and mitigate the problem of reward sparsity.

Summary by alphaXiv.