Must chance outcomes be equally likely?

No. Each outcome has its own probability, and evaluation uses those weights.

Can an action change a chance node's probabilities?

Yes. The distribution can depend on earlier decisions or other variables in the model.

Does a higher expected utility mean every result is better?

No. An action can have a higher average utility while retaining an unfavorable possible outcome.

AI Glossary / Graphs and symbolic AI / Intermediate

Chance nodes

Definition

Chance nodes represent uncertain outcomes in a decision tree, game tree, or decision network. Their alternatives are governed by probabilities rather than selected directly by the decision maker. When evaluating expected utility in a tree, a chance node combines the values of its possible outcomes using their probabilities as weights.

What it means

A decision node asks which action to choose. A chance node asks what may happen, given the relevant situation and action. A third kind of node may represent an opposing decision maker, whose behavior requires its own model. Keeping these roles distinct prevents a planner from treating favorable uncertain events as though it could simply choose them.
The uncertainty can concern a future random event or an unknown condition represented probabilistically. Drawing a chance node does not establish that its probabilities are correct. It records the modeler's representation of uncertainty, which must be supported and updated as appropriate.

How it works

In a tree, give the outgoing outcomes probabilities that form a valid distribution for that node's context. Calculate the value of each continuation, multiply it by the corresponding probability, and add the results. A decision node can then compare the resulting action values.
In a stochastic game tree, expectiminimax combines maximizing, minimizing, and chance evaluations. In a decision network, chance nodes represent random variables with conditional distributions, and inference supports expected-utility calculations. These representations differ in structure, but both distinguish uncertain events from controllable choices. An ordinary arithmetic average is appropriate only when the relevant outcomes are equally likely.

Three practical examples

A game-playing agent evaluates the possible results of a die roll before choosing a later move.
An operations planner represents whether a delivery arrives within a needed window after a dispatch decision.
A decision-support model represents an uncertain equipment condition that influences the consequences of alternative actions.

Think of it this way

Think of a fork in a plan where the next branch is determined by an event rather than by the planner. The planner can prepare for the branches but cannot choose the favorable one merely by wishing. The analogy is limited because some actions can influence the probabilities; uncontrollable outcome selection does not imply that every action leaves the distribution unchanged.

Worked example

A fictional decision offers two actions measured in utility points. Action A leads to a chance node with a 0.70 probability of 10 points and a 0.30 probability of -4 points. Its expected utility is 0.70 times 10 plus 0.30 times -4, or 5.8. Action B gives 5 points for certain.
An expected-utility maximizer selects A under these stated values and probabilities. Selecting 10 as A's value would incorrectly treat the chance node as a maximizing decision node. Taking the unweighted average of 10 and -4 gives 3, also incorrect because the outcomes are not equally likely. If the probability of the favorable A outcome is p, its expected utility is 14p minus 4. It exceeds 5 only when p is greater than 9 divided by 14, about 0.643. This sensitivity calculation shows why an uncertain probability estimate near the threshold deserves attention.

Illustrative case study

At fictional Fir Dispatch, planners compare two delivery arrangements using a decision tree. They separate the choice of arrangement from the uncertain arrival outcome and document different arrival probabilities for the alternatives. Reviewers check that the probabilities describe the intended operating conditions.
When conditions change, the team revisits the chance model rather than treating the earlier preferred action as permanently best. This fictional case illustrates the role of uncertainty in action evaluation. It does not claim that a favorable expected result guarantees a successful individual delivery.

Common misconceptions and mistakes

A chance node is not a decision node: selecting its best branch overstates control. It is also not automatically a worst-case node. Worst-case planning and expected-utility planning use different evaluation rules.
Probabilities must describe mutually exclusive, collectively exhaustive outcomes at the modeled level, or the representation must account for omissions and overlaps. Do not average raw monetary outcomes while claiming to represent a nonlinear utility preference. Finally, a probability model for an adversary is an assumption about behavior; a strategically choosing opponent should not be treated as a fair random device without justification.

Business use

Chance nodes help businesses make uncertainty explicit in decision diagrams. They encourage teams to identify which quantities are choices, which are uncertain, and which evidence supports the probabilities. Useful reviews examine sensitivity, consequences of unfavorable branches, and any hard conditions that expected utility does not replace. The model can guide a decision, but it should preserve the distinction between a recommended action and a guaranteed outcome.

Check your understanding

A chance node has outcomes worth 20 and 0 with probabilities 0.10 and 0.90. Is its expected value 10?

Answer: No. The probability-weighted value is 2. Ten would be the unweighted average and would assume equal probabilities.

Related terms

Learn next

Draw the worked decision with square nodes for choices and circles for uncertain outcomes. Calculate values backward from the leaves, labeling every probability. Then vary one probability and identify when the preferred action changes. This tests whether the recommendation is stable under plausible uncertainty.

Sources

Sources: Russell and Norvig, original AIMA game-playing slides, stochastic games and expectiminimax: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter06.pdf . Original AIMA utility and decision-making slides: https://aima.cs.berkeley.edu/slides-pdf/chapter16.pdf .

Previous: Chambolle - Pock algorithm Next: Channel