Learning Approximate Stochastic Transition Models

TL;DR

This paper introduces a modified GAN-based approach for learning stochastic transition models in reinforcement learning, capable of generalizing to new states and capturing randomness in state transitions.

Contribution

It proposes a novel modification to GAN loss functions that enables effective learning of stochastic transition models, overcoming limitations of existing methods.

Findings

Modified GAN loss functions improve stochastic transition modeling.

The approach generalizes well to unseen states.

It outperforms traditional GANs in capturing stochastic dynamics.

Abstract

We examine the problem of learning mappings from state to state, suitable for use in a model-based reinforcement-learning setting, that simultaneously generalize to novel states and can capture stochastic transitions. We show that currently popular generative adversarial networks struggle to learn these stochastic transition models but a modification to their loss functions results in a powerful learning algorithm for this class of problems.