Implements Generative Flow Networks to sample diverse candidates proportionally to their reward functions.
The GFlowNet skill empowers Claude to design and implement Generative Flow Networks (GFlowNets), a machine learning paradigm developed by Yoshua Bengio that bridges the gap between Reinforcement Learning and MCMC sampling. Unlike traditional RL that converges on a single optimal solution, GFlowNets learn to sample diverse high-reward candidates—such as molecular structures or causal graphs—with probabilities proportional to their reward. This skill is ideal for researchers and engineers working on drug discovery, causal inference, and combinatorial optimization where finding a variety of high-quality solutions is more critical than finding a single global maximum.
Key Features
012 GitHub stars
02Implementation of Trajectory Balance (TB) and Detailed Balance (DB) training objectives
03Tools for Bayesian structure learning and causal discovery of DAGs
04Integration with scientific computing libraries for topological and chemical analysis
05Boilerplates for combinatorial optimization and set generation
06Support for diverse candidate generation in molecule design and drug discovery
Use Cases
01Solving combinatorial optimization problems where exploration of multi-modal reward landscapes is required
02Performing causal discovery to infer complex relationships between variables from observational data
03Discovering diverse drug candidates with high binding affinity and synthesizability
