QGFN: Controllable Greediness with Action Values

Authors: Elaine Lau, Stephen Lu, Ling Pan, Doina Precup, Emmanuel Bengio

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the proposed methods on 5 standard tasks used in prior GFN works: the fragment-based molecular design task introduced by Bengio et al. [2021a], 2 RNA design tasks introduced by Sinai et al. [2020], a small molecule design task based on QM9 [Jain et al., 2023], as well as a bit sequence task from Malkin et al. [2022a], Shen et al. [2023]. The proposed method outperforms strong baselines, achieving high average rewards and discovering modes more efficiently, sometimes by a large margin.
Researcher Affiliation Collaboration Elaine Lau1 2 Stephen Zhewen Lu2 Ling Pan1 5 Doina Precup1 2 3 Emmanuel Bengio4 1Mila Québec AI Institute 2Mc Gill University 3Google Deepmind 4Valence Labs 5Hong Kong University of Science and Technology
Pseudocode Yes Algorithm 1 QGFN: Full training algorithm details
Open Source Code Yes Source code available at: https://github.com/yunglau/QGFN/
Open Datasets Yes We evaluate the proposed methods on 5 standard tasks used in prior GFN works: the fragment-based molecular design task introduced by Bengio et al. [2021a], 2 RNA design tasks introduced by Sinai et al. [2020], a small molecule design task based on QM9 [Jain et al., 2023], as well as a bit sequence task from Malkin et al. [2022a], Shen et al. [2023].
Dataset Splits Yes We evaluate the proposed methods on 5 standard tasks used in prior GFN works: the fragment-based molecular design task introduced by Bengio et al. [2021a], 2 RNA design tasks introduced by Sinai et al. [2020], a small molecule design task based on QM9 [Jain et al., 2023], as well as a bit sequence task from Malkin et al. [2022a], Shen et al. [2023].
Hardware Specification Yes All of our experiments were conducted using A100 and V100 GPUs.
Software Dependencies No The paper mentions software and libraries like 'RDKit' and 'Vienna RNA', but it does not provide specific version numbers for any of these software dependencies.
Experiment Setup Yes Parameter Value Batch size 64 Number of steps 10,000 Optimizer Adam Number of Layers 4 Hidden Dim. Size 128 Number of Heads 2 Positional Embeddings Rotary Reward scaling β in Rβ 32 Learning rate 1 10 4 Z Learning rate 1 10 3