Dictionary of Applied Machine Learning
Updated on 2026-09-18
Typeset PDF version — the authoritative form of this entry
A policy in reinforcement learning (RL) is a function that specifies which action $\arm_{\timeidx}$ to take in a Markov decision process (MDP) when the current state is $\state_{\timeidx}$. A policy is often stochastic: it defines a conditional probability distribution $\probdist^{(\arm \mid \state)}$ over the actions given the state, and an action is drawn from it. A policy is the hypothesis of RL: it uses features derived from the state to predict the best next action, and it is learned to maximize the expected cumulative reward. The value function of a policy is the expected cumulative reward obtained by following it from a given state.
A policy is a function that specifies how the next action $\arm_{\timeidx}$ in a Markov decision process (MDP) is chosen when the current state is $\state_{\timeidx}$. Typically, a policy is stochastic, meaning that it defines a conditional probability distribution $\probdist^{(\arm \mid \state)}$ over the actions for a given current state. A policy can also be viewed as a hypothesis that uses features derived from the current state to predict the best next action (Sutton and Barto, 2018).
See also: action, Markov decision process, state.
@misc{dictml_policy,
author = {Jung, Alexander and Olioumtsevits, Konstantina and Schnoor, Ekkehard},
title = {policy},
howpublished = {Dictionary of Applied Machine Learning (course edition)},
year = {2026},
doi = {10.5281/zenodo.21569296},
note = {ISBN 978-952-64-3013-3, CC BY 4.0, retrieved 2026-09-19},
url = {https://dictionaryofml.org/terms/policy.html}
}