2
$\begingroup$

I am reading the Yoshua Bengio et al, Flow Network based Generative Models for Non-Iterative Diverse Candidate Generation. After read this paper, I wonder how to establish a GFlowNet if I don't know all the States? That is, I only have goals, such as maximum likelihood estimation or some generation models similar to MCMC. How can I initialize the GFlowNet nodes when I don't know all the states in advance?

$\endgroup$
1
  • $\begingroup$ This paper might provide you with some intuition: Generative Flow Networks for Discrete Probabilistic Modeling arxiv.org/pdf/2202.01361.pdf $\endgroup$
    – jzin
    May 13, 2022 at 8:27

1 Answer 1

2
$\begingroup$

As in deep reinforcement learning, one does not need to literally hold the state space in memory explicitly (i.e. we don't need to "know all the states"). Rather, what is used is a function approximator that maps a description of a state to a prediction. In DRL that would be for example $Q_\theta(s,a)$, or $\pi_\theta(a|s)$.

In this paper, the flow $F_\theta(s,a)$ is what is being approximated, with a deep neural network. All that is needed is thus a description of the current state $s$ that can be given to a neural network (e.g. an image). In the paper, a graph is used as an input to a graph neural network. These graphs are not to be confused with the state space DAG, which in this work is a graph of (molecular) graphs, where an edge represents a graph edit.

Disclaimer, I am the first author of this paper.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.