liveBook · Manning

Policy

As explained in Grokking Bayes

A policy is a rule or strategy that selects actions based on the available information or state of a decision problem.

A policy is an available action or strategy for improving an outcome under uncertainty. Its value depends on possible effects, implementation costs, and the information available when decisions are made. section 12.7

Choosing between two policies

A government team can fund Policy A, a low-cost training program with modest impact; fund Policy B, an expensive program with a small chance of a much larger impact; or do nothing. Policy A has expected net utility −1 + 0.5 × 10 + 0.5 × 5 = 6.5. Policy B has −2 + 0.1 × 30 + 0.9 × 0 = 1. Because both policies outperform doing nothing, Policy A is the best one-step choice. section 12.7

With multiple decision steps, a policy is evaluated by its consequences across the whole sequence. After observing an outcome, later actions can change: terminate, continue with another policy, or choose the alternative that now has greater expected utility. section 12.8

A policy can support two competing decision goals:

  • Exploit the action with the highest immediate expected utility, such as choosing the stable Policy A. section 12.9
  • Explore an uncertain action, such as Policy B, when its possible future payoff could improve later decisions. section 12.9
  • Balance both goals by propagating terminal utilities backward through a search tree and selecting the action with the greatest adaptive expected utility. section 12.8 section 12.9
Table 12.3 Utility of two available policies under possible scenarios and the corresponding probabilities
Option
Policy A
Policy B
Doing nothing
Cost
1
2
0
Utility
10
5
30
0
0
Probability
0.5
0.5
0.1
0.9
1
Option
Cost
Utility
Probability
Policy A
1
10
0.5
5
0.5
Policy B
2
30
0.1
0
0.9
Doing nothing
0
0
1

Compare each policy's cost, possible utility, and probability before calculating expected net utility. section 12.7

How Grokking Bayes covers it

How deep it goes
Explains one-step expected utility and multistep utility propagation through search trees, without implementation code.
Examples
Uses government training policies, costs, possible effects, probabilities, and restaurant choices to illustrate sequential decisions.
What it assumes
Assumes familiarity with probabilities, costs, utilities, actions, and decision trees while building sequential reasoning from simple comparisons.
What you can do after
Evaluate policy alternatives by calculating expected utility and selecting adaptive follow-up actions in a decision tree.
How it connects
Connects policies to Bayesian decision theory, expected utility, search trees, adaptive decisions, and exploration–exploitation.

Also explained in

Compare how 21 books explain policy →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Grokking Bayes. The section links point to the parts of the book it draws on.