Finetuned Decision Transformer with Tree Search for Logic Synthesis Optimization

Logic synthesis (LS) involves the conversion of high-level circuit descriptions into gate-level netlists. However, identifying the optimal primitive sequences (PS) for this transformation is a formidable challenge due to the expansive design space. While machine learning has been employed to address this issue, existing methods often necessitate extensive training or result in high computational costs. GPT-LS revolutionized LS optimization by reframing it as a sequence generation problem and leveraging a pre-trained transformer with offline reinforcement learning to generate PS effectively. In this work, we integrate tree search and model fine-tuning to enhance the transferability and performance of GPT-LS, particularly in unseen cases, which significantly improves efficiency and effectiveness with minimal runtime increase, establishing a new standard in LS optimization. Experimental results demonstrate that, after just a few rounds of fine-tuning, it outperforms all previous methods, including state-of-the-art Bayesian optimization techniques and long-running random searches.