Energy Based GFlowNets

Energy-based Generative Flow Networks (GFlowNets) are a class of generative models designed to efficiently sample complex objects from unnormalized probability distributions, often defined by a reward function. Current research focuses on improving GFlowNet training through enhanced loss functions, novel exploration strategies (like adapted metadynamics and Monte Carlo Tree Search), and architectural modifications such as bifurcated networks and incorporating symmetry or local search. These advancements are significantly impacting diverse fields, enabling more efficient exploration of vast combinatorial spaces in applications ranging from drug discovery and materials science to sensor selection and combinatorial optimization.