A Distributional Analogue to the Successor Representation

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wiltzer, Harley, Farebrother, Jesse, Gretton, Arthur, Tang, Yunhao, Barreto, André, Dabney, Will, Bellemare, Marc G., Rowland, Mark
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909209668354048
author Wiltzer, Harley
Farebrother, Jesse
Gretton, Arthur
Tang, Yunhao
Barreto, André
Dabney, Will
Bellemare, Marc G.
Rowland, Mark
author_facet Wiltzer, Harley
Farebrother, Jesse
Gretton, Arthur
Tang, Yunhao
Barreto, André
Dabney, Will
Bellemare, Marc G.
Rowland, Mark
contents This paper contributes a new approach for distributional reinforcement learning which elucidates a clean separation of transition structure and reward in the learning process. Analogous to how the successor representation (SR) describes the expected consequences of behaving according to a given policy, our distributional successor measure (SM) describes the distributional consequences of this behaviour. We formulate the distributional SM as a distribution over distributions and provide theory connecting it with distributional and model-based reinforcement learning. Moreover, we propose an algorithm that learns the distributional SM from data by minimizing a two-level maximum mean discrepancy. Key to our method are a number of algorithmic techniques that are independently valuable for learning generative models of state. As an illustration of the usefulness of the distributional SM, we show that it enables zero-shot risk-sensitive policy evaluation in a way that was not previously possible.
format Preprint
id arxiv_https___arxiv_org_abs_2402_08530
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle A Distributional Analogue to the Successor Representation
Wiltzer, Harley
Farebrother, Jesse
Gretton, Arthur
Tang, Yunhao
Barreto, André
Dabney, Will
Bellemare, Marc G.
Rowland, Mark
Machine Learning
Artificial Intelligence
This paper contributes a new approach for distributional reinforcement learning which elucidates a clean separation of transition structure and reward in the learning process. Analogous to how the successor representation (SR) describes the expected consequences of behaving according to a given policy, our distributional successor measure (SM) describes the distributional consequences of this behaviour. We formulate the distributional SM as a distribution over distributions and provide theory connecting it with distributional and model-based reinforcement learning. Moreover, we propose an algorithm that learns the distributional SM from data by minimizing a two-level maximum mean discrepancy. Key to our method are a number of algorithmic techniques that are independently valuable for learning generative models of state. As an illustration of the usefulness of the distributional SM, we show that it enables zero-shot risk-sensitive policy evaluation in a way that was not previously possible.
title A Distributional Analogue to the Successor Representation
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2402.08530