# Flow Network based Generative Models for Non-Iterative Diverse Candidate Generation

**Authors:** Emmanuel Bengio, Moksh Jain, Maksym Korablyov, Doina Precup, Yoshua Bengio  
**arXiv:** [2106.04399](https://arxiv.org/abs/2106.04399)  
**Date:** 2021-06-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.04399v1/

## Key points

- The paper addresses the problem of learning a stochastic policy for generating objects from a sequence of actions.
- GFlowNet is proposed as a novel approach based on Temporal Difference learning and flow networks.
- GFlowNet aims to sample diverse sets of high-return solutions, which is useful in scenarios like black-box function optimization or molecule design.
- Unlike traditional return maximization methods, GFlowNet allows for fast generation by amortizing the cost of search during training.
- GFlowNet treats the generative process as a flow network, enabling it to handle cases where different trajectories can yield the same final state.
- The proposed method has potential applications in various domains, including molecule synthesis and black-box function optimization.

## Summary

, , , , 

This paper addresses the problem of learning a stochastic policy for generating objects, such as molecular graphs, from a sequence of actions. The goal is to generate objects with a probability proportional to a given positive reward for that object. The authors propose GFlowNet, a novel approach based on Temporal Difference learning and the concept of flow networks. GFlowNet aims to sample diverse sets of high-return solutions, which is particularly useful in scenarios like black-box function optimization or molecule design. Unlike traditional return maximization methods that converge to a single sequence, GFlowNet allows for fast generation by amortizing the cost of search during training. This paper introduces GFlowNet as an alternative approach to convert an energy function into a fast generative model. A novel approach based on Temporal Difference learning and flow networks for generating objects with diverse sets of high-return solutions. The proposed method treats the generative process as a flow network, enabling it to handle cases where different trajectories can yield the same final state. GFlowNet aims to sample diverse sets of high-return solutions, providing an advantage in applications like drug discovery. The proposed method has potential applications in various domains, including molecule synthesis and black-box function optimization.