Hierarchical Cooperative GFlowNets coordinate sampler exploration
Original title: Hierarchical Cooperative GFlowNets
A shared classifier partitions the search space so lower-ranked models cover regions higher-ranked ones miss.
The paper proposes Hierarchical Cooperative GFlowNets to stop independently trained GFlowNets from collapsing onto the same high-reward regions. The highest-ranked model may explore the full space, while lower-ranked models are pushed toward uncovered regions without hand-crafted boundaries. A shared classifier partitions the space while the GFlowNets stay independently parameterized and can train asynchronously. The authors derive a normalizer-weighted ensemble rule and report better mode coverage and target approximation on multimodal, sequence, combinatorial, and biological sequence-design tasks.
Unconfirmed: Recovery of the target distribution is proved only under the paper's stated assumptions; gains are author-reported.
- Independent parallel GFlowNets can still collapse onto the same high-reward areas.
- Lower-ranked models are progressively steered away from regions covered by higher-ranked ones.
- The ensemble rule is proved to recover the target distribution under stated assumptions.
- Experiments include sequence generation, combinatorial optimization, and biological sequence design.
From the original article
1 min read at arxiv.orgGenerative Flow Networks (GFlowNets) are trained to sample objects from compositional spaces with probability proportional to a reward function, but ineffective exploration can cause learned samplers to concentrate on a few high-reward regions. An intuitive remedy is to train multiple GFlowNets independently in parallel and combine their samples.
By Pedro Dall'Antonia · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.