Controlling Exploration-Exploitation in GFlowNets via Markov Chain Perspectives
This work addresses the exploration-exploitation problem in GFlowNets for generative modeling, offering a tunable method to improve mode discovery, though it is incremental as it builds on existing GFlowNet frameworks.
The paper tackled the limitation of fixed exploration-exploitation trade-offs in GFlowNets by linking them to Markov chain reversibility, proposing α-GFNs with a tunable parameter that achieved up to a 10× increase in discovered modes across benchmarks.
Generative Flow Network (GFlowNet) objectives implicitly fix an equal mixing of forward and backward policies, potentially constraining the exploration-exploitation trade-off during training. By further exploring the link between GFlowNets and Markov chains, we establish an equivalence between GFlowNet objectives and Markov chain reversibility, thereby revealing the origin of such constraints, and provide a framework for adapting Markov chain properties to GFlowNets. Building on these theoretical findings, we propose $α$-GFNs, which generalize the mixing via a tunable parameter $α$. This generalization enables direct control over exploration-exploitation dynamics to enhance mode discovery capabilities, while ensuring convergence to unique flows. Across various benchmarks, including Set, Bit Sequence, and Molecule Generation, $α$-GFN objectives consistently outperform previous GFlowNet objectives, achieving up to a $10 \times$ increase in the number of discovered modes.