← Back to selected work
02 / Machine learning research
Make every
guess count.
When you can’t afford to evaluate every possibility, choosing what
to try next becomes the real problem.
PyTorchGFlowNetsGaussian processesActive learning
A familiar game.
A different question.
Think of building a word, one letter at a time. There are many
possible sequences, but only some are useful. Now put a limit on
how often you can ask for a score.
GFN-AL uses a Scrabble environment to compare ways of spending
that evaluation budget: active learning, a directly trained
GFlowNet, and a hybrid of the two.
A WORLD OF POSSIBLE ANSWERS
S1E1A1R1C3H4
ONE LIMITED BUDGET TO FIND THEM.
The experiment
Three routes through the same search.
Each method can propose candidates, ask the oracle for their
scores, and use what it learns to guide the next step. The scarce
resource is the number of oracle queries.
-
Active learning: a Gaussian-process surrogate
predicts candidate quality. An upper-confidence-bound
acquisition rule balances promising predictions with
uncertainty.
-
Direct GFlowNet: learns to generate candidates
using rewards from the oracle.
-
The hybrid: brings surrogate-guided learning
and GFlowNet sampling into the same loop.
Under the hood
Ask, learn, propose, repeat.
01Propose a pool of candidate sequences
↓
02Rank candidates using predictions and uncertainty
↓
03Spend part of the query budget on selected candidates
↓
04Update the surrogate and repeat within the budget
The code separates the oracle, acquisition rule, surrogate, and
training loops. Configuration controls the budget, candidate pool,
batch size, and sampling strategy.
What to measure
Finding one answer isn’t the whole story.
A useful comparison considers score, valid candidates, diversity,
and how these change as queries are consumed. A method that
repeatedly finds the same good answer behaves differently from one
that discovers a varied set of useful answers.
The repository contains three experiment entry points and smoke
tests. This page describes the experiment design; it does not
present an independently reproduced benchmark or a winner.
Built on
A shared research foundation.
The Scrabble environment and scorer come from
the upstream GFlowNet library. GFN-AL builds its comparison and training workflows around that
dependency. The upstream library is credited separately from this
project.