Peter Stone's Selected Publications

Classified by TopicClassified by Publication TypeSorted by DateSorted by First Author Last NameClassified by Funding Source


Proto Successor Measure: Representing the Behavior Space of an RL Agent

Proto Successor Measure: Representing the Behavior Space of an RL Agent.
Siddhant Agarwal, Harshit Sikchi, Peter Stone, and Amy Zhang.
In International Conference on Machine Learning, June 2025.

Download

[PDF]911.1kB  

Abstract

Having explored an environment, intelligent agents should be able to transfertheir knowledge to most downstream tasks within that environment withoutadditional interactions. Referred to as "zero-shot learning", this abilityremains elusive for general-purpose reinforcement learning algorithms. Whilerecent works have attempted to produce zero-shot RL agents, they make assumptionsabout the nature of the tasks or the structure of the MDP. We present ProtoSuccessor Measure: the basis set for all possible behaviors of a ReinforcementLearning Agent in a dynamical system. We prove that any possible behavior(represented using visitation distributions) can be represented using an affinecombination of these policy-independent basis functions. Given a reward functionat test time, we simply need to find the right set of linear weights to combinethese bases corresponding to the optimal policy. We derive a practical algorithmto learn these basis functions using reward-free interaction data from theenvironment and show that our approach can produce the near-optimal policy attest time for any given reward function without additional environmentalinteractions.

BibTeX Entry

@InProceedings{siddhant_agarwal_psm,
  author   = {Siddhant Agarwal and Harshit Sikchi and Peter Stone and Amy Zhang},
  title    = {Proto Successor Measure: Representing the Behavior Space of an RL Agent},
  booktitle = {International Conference on Machine Learning},
  year     = {2025},
  month    = {June},
  location = {Vancouver, Canada},
  abstract = {Having explored an environment, intelligent agents should be able to transfer
their knowledge to most downstream tasks within that environment without
additional interactions. Referred to as "zero-shot learning", this ability
remains elusive for general-purpose reinforcement learning algorithms. While
recent works have attempted to produce zero-shot RL agents, they make assumptions
about the nature of the tasks or the structure of the MDP. We present Proto
Successor Measure: the basis set for all possible behaviors of a Reinforcement
Learning Agent in a dynamical system. We prove that any possible behavior
(represented using visitation distributions) can be represented using an affine
combination of these policy-independent basis functions. Given a reward function
at test time, we simply need to find the right set of linear weights to combine
these bases corresponding to the optimal policy. We derive a practical algorithm
to learn these basis functions using reward-free interaction data from the
environment and show that our approach can produce the near-optimal policy at
test time for any given reward function without additional environmental
interactions.
  },
}

Generated by bib2html.pl (written by Patrick Riley ) on Mon Sep 29, 2025 19:22:37