Policy

Scope: A decision rule in the partially observable planning setting of Kaelbling, Littman, and Cassandra (1998) ¹; “policy” in governance or natural-language instructions has different meanings.

An agent function describes behavior over percept histories. ² The belief-state policy below instead uses a summary of the information available in the specified POMDP. ¹ Their input descriptions differ, so the terms are not automatically interchangeable.

In the POMDP model

A policy determines how to choose an Action from information available at decision time. Since the true state is hidden, a rule that chooses an action from the true state cannot in general be implemented by this agent. A belief-state policy instead uses a distribution over possible states: symbolically, $\pi (b)=a$ means the rule $\pi$ selects action $a$ at Belief state $b$. This shorthand is for a deterministic, stationary rule; history-dependent or stochastic rules are also possible, and the shorthand is not their universal definition. Planning evaluates such choices with the task’s rewards and assumptions about future transitions and observations.

Boundary example

“If certain of an obstacle, turn; otherwise inspect” can be a decision rule only once “certain,” “turn,” and “inspect” are specified within a model. A one-off turn after a noisy sensor reading is an action, not the whole policy. A document labeled “safety policy” is not automatically the mathematical action-selection rule considered here.

Relations

• Maps from: a Belief state can summarize action–observation history for planning in this setting.
• Specifies: a deterministic policy selects an Action; a stochastic policy specifies a probability distribution over actions.
• Responds to: observations affect later choices through the updated belief state.
• Contrast: Retrieval-augmented generation describes a retrieval-conditioned generation architecture, not necessarily a policy optimizing sequential reward.

Footnotes

1. Kaelbling, Leslie Pack, Michael L. Littman, and Anthony R. Cassandra. “Planning and Acting in Partially Observable Stochastic Domains.” 1998. Artificial Intelligence 101 (1–2): 99–134. https://doi.org/10.1016/S0004-3702(98)00023-X. Author-hosted paper PDF. ↩ ↩²

2. Russell, Stuart J., and Peter Norvig. Artificial Intelligence: A Modern Approach. Fourth edition, Global Edition. Pearson Education Limited, 2022. Chapter 2, §2.1, pp. 54–56. Official edition page. ↩