Can perfect information have zero value?

Yes, for example when an action remains optimal for every possible observation.

Must the information reveal everything?

No. The observed variable must be specified; perfect observation of one variable can leave other uncertainty.

Is the value known only after observing?

No. The expected value is calculated beforehand by averaging over possible observations.

AI Glossary / Graphs and symbolic AI / Advanced

Value of perfect information

Definition

The value of perfect information is the expected improvement in the best achievable utility from learning an uncertain variable's exact value before making a decision. It compares the optimal informed decision with the optimal decision using existing information. The usual value excludes observation costs, which must be considered separately.

What it means

Information matters because it can improve a choice, not simply because it reduces uncertainty. A fact may be surprising or highly informative while leaving the best available action unchanged. In that case, its decision value can be zero.
Specify what becomes perfectly known. Exact knowledge of a forecast is not exact knowledge of the weather. Observing one variable may leave other uncertainty unresolved. Full perfect information about all decision-relevant uncertainty provides a different benchmark from perfect observation of one selected variable.

How it works

First compute the highest expected utility available under current evidence. Then consider every possible value of the variable to be observed. For each value, update beliefs, choose the best action, and calculate its conditional expected utility. Average those optimized values using the probabilities of the observations, and subtract the original optimum.
The order matters: with information, the action can depend on what is observed. Choosing one action before averaging describes the uninformed situation instead.
With costless information, unchanged opportunities, and permission to ignore it, the expected value cannot be negative. A particular observation can bring bad news; the guarantee concerns the average value of observing. Delays, fees, and other consequences of obtaining information change the practical calculation.

Three practical examples

A warehouse compares stocking plans before uncertain demand is realized. Knowing demand in advance would allow a different quantity choice in each demand state, providing a benchmark for evaluating forecasting work.
A maintenance team evaluates learning a component's exact condition before choosing an intervention. If the same intervention is best in every possible condition, perfect condition information adds no decision value within that model.
An event organizer considers buying a predictive report. Its value depends on accuracy and arrival before commitments become irreversible.

Think of it this way

Imagine choosing equipment before opening a sealed envelope that reveals tomorrow's requirement. Perfect information means opening the envelope first and then choosing. Its value is the expected benefit of adapting that choice, compared with making the best choice while the envelope remains closed.

Worked example

A warehouse faces high demand with probability 0.3 and low demand with probability 0.7. A large plan has utility 100 under high demand and -40 under low demand. A small plan has utility 50 under high demand and 20 under low demand.
Without new information, the large plan scores 0.3 × 100 + 0.7 × (-40) = 2. The small plan scores 0.3 × 50 + 0.7 × 20 = 29, so the best uninformed choice is small.
With exact demand information, choose large when demand is high and small when it is low. Expected utility becomes 0.3 × 100 + 0.7 × 20 = 44. Perfect information is therefore worth 44-29=15 utility units before costs.
Now consider a noisy signal that reports high with probability 0.8 when demand is high, and low with probability 0.8 when demand is low. The joint probabilities for actual high/low demand and a high signal are 0.24 and 0.14; for a low signal they are 0.06 and 0.56.
The optimal signal-based rule chooses large after a high signal and small after a low signal. Its expected utility is 0.24 × 100 + 0.14 × (-40) + 0.06 × 50 + 0.56 × 20 = 32.6. This signal is worth only 3.6 units above the original 29, below the perfect-information benchmark of 15.

Illustrative case study

In a fictional operations team using the worked-example model, a forecast supplier emphasizes the value of eliminating uncertainty. Reviewers distinguish exact observation from the noisy signal.
Assume obtaining the noisy signal imposes an additive cost of 2 utility units, with no other effects. Its net expected utility is 32.6-2=30.6, exceeding the uninformed optimum of 29 by 1.6. That supports acquisition under these specific assumptions. A cost of 5 would instead make this signal unattractive, even though 5 is below the perfect-information value of 15.
The team also checks delivery timing and whether the stocking decision can still change. It uses the ideal benchmark to screen possibilities, then evaluates the actual information product. The benchmark alone does not establish that any particular forecast is worth its price.

Common misconceptions and mistakes

“Perfect information guarantees a good outcome.” It improves the choice under uncertainty; the observed state may still be unfavorable.
“More accurate information always has positive decision value.” It can leave the optimal decision unchanged.
“The perfect-information value is what every forecast is worth.” A noisy or incomplete signal generally provides less value under the same decision setup.
“Utility units are automatically currency.” Monetary willingness to pay needs an appropriate treatment of money, utility, and costs.
“Values for two observations can always be added.” Information may overlap or become useful only in combination, so joint value requires a joint calculation.

Business use

Use perfect information as an upper benchmark when considering additional data, inspection, or prediction about the same uncertainty. If even exact knowledge cannot justify the acquisition burden, a less informative signal with no additional benefits cannot do better under the same model.
For a real purchase, evaluate the actual signal's reliability, timing, and consequences. Include the best existing decision as the baseline; comparing with an arbitrary weak action exaggerates information value. Review the utility model and probabilities, since the result depends on both. If information changes what actions are available or has benefits beyond the modeled decision, expand the model before applying the benchmark mechanically.

Check your understanding

What is the best uninformed utility in the example?

Answer: 29 from the small plan.

Why is perfect information worth 15?

Answer: It raises optimized expected utility from 29 to 44.

Why is the noisy signal worth less?

Answer: Its errors sometimes cause an unsuitable plan to be chosen.

Related terms

Expected utility
Decision network
Decision analysis
Optimal policy
value of information
value of sample information
Bayesian updating
information acquisition

Learn next

Calculate the value of an imperfect observation by updating beliefs for each possible signal. Then compare gross decision value with acquisition costs and decision timing.

Sources

University of California, Berkeley, CS 188, The Value of Perfect Information: https://inst.eecs.berkeley.edu/~cs188/textbook/vpis/vpi.html Russell and Norvig, AIMA, Rational Decisions lecture slides: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter16.pdf AIMA glossary, value of perfect information: https://github.com/aimacode/aima-glossary

Previous: Value iteration Next: Value sensitive design