Reproduce “Goal-conditioned GFlowNets for Controllable Multi-Objective Molecular Design” on OpenResearch

About this paper

Julien Roy, Pierre-Luc Bacon, Christopher Pal, Emmanuel Bengio

Goal-conditioned GFlowNets (GC-GFNs) were developed for molecular design, enabling the generation of molecules whose properties align with explicit objective goals. This approach achieves superior controllability and more uniform exploration of complex multi-objective landscapes compared to preference-conditioned methods, particularly with concave Pareto fronts and increasing numbers of objectives.

Key insights

• Hard goal conditioning is critical for achieving fine-grained control and uniform exploration of multi-objective spaces, particularly when dealing with non-convex or concave Pareto fronts, outperforming soft preference conditioning.
• The combination of a replay buffer and Hindsight Experience Replay (HER) is essential for stable learning in goal-conditioned, sparse-reward generative tasks.
• An adaptive goal-sampling mechanism, like the Tabular Goal-Sampler, significantly improves sample efficiency and scalability when dealing with increasing numbers of objectives by focusing on feasible regions.

Summary by alphaXiv.