A Graph Enhanced Symbolic Discovery Framework For Efficient Logic Optimization
Authors: Yinqi Bai, Jie Wang, Lei Chen, Zhihai Wang, Yufei Kuang, Mingxuan Yuan, Jianye HAO, Feng Wu
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on three challenging circuit benchmarks show that the interpretable symbolic functions learned by CMO outperform previous state-of-the-art (SOTA) GPU-based and human-designed approaches in terms of inference efficiency and generalization capability. Moreover, we integrate CMO with the Mfs2 heuristic one of the most time-consuming LO heuristics. The empirical results demonstrate that CMO significantly improves its efficiency while keeping comparable optimization performance when executed on a CPU-based machine, achieving up to 2.5 faster runtime. |
| Researcher Affiliation | Collaboration | 1Mo E Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China 2 Noah s Ark Lab, Huawei Technologies 3 College of Intelligence and Computing, Tianjin University |
| Pseudocode | Yes | 1. Algorithm. We provide the architecture and pseudocode of our approach in Section 4 and Appendix E.5. Moreover, we will make our source code publicly available once the paper is accepted for publication. |
| Open Source Code | Yes | 2. Source Code. To facilitate the evaluation process and support a thorough review, we have released our source code at the following link: https://gitee.com/yinqi-bai/ cmo.git. |
| Open Datasets | Yes | Benchmarks We evaluate CMO on two widely-used public benchmarks, EPFL (Amar u et al., 2015) and IWLS (Albrecht, 2005), and one industrial benchmark from an anonymous semiconductor company. |
| Dataset Splits | Yes | Inspired by the leave-one-domain-out cross-validation strategy commonly used in previous literature (Wang et al., 2022a), we design twelve leave-one-out datasets for evaluation. Specifically, given a benchmark, we construct a dataset by setting one circuit as the testing dataset, and the other circuits as the training dataset. Please refer to D.2 for more details. |
| Hardware Specification | Yes | Experiments are performed on a single machine that contains 32 Intel Xeon R E5-2667 v4 CPUs. |
| Software Dependencies | No | Throughout all experiments, we use ABC (Brayton et al., 2010) as the backend LO framework. ABC is a state-of-the-art open-source LO framework and is widely used in research of machine learning for LO (Pasandi et al., 2023). |
| Experiment Setup | Yes | In our implementation, we configured the weight factor λ between Lteacher and Lstudent as 0.5 to balance the objective functions. We averaged the results over three different random seeds during training. Additionally, a regularization factor ρ = 0.99 was employed to penalize excessive expression length, encouraging more compact and interpretable solutions. The training process utilized Monte Carlo Tree Search (MCTS), running 10,000 episodes per iteration across 20 iterations. After training, the best-performing model on the training set from all 20 iterations was selected, ensuring both stability and optimal performance. |