Must an action cost be money?

No. It can represent time, energy, action count, or another clearly defined quantity.

Is a path's cost always its number of steps?

Only under a unit-cost model. Other functions assign different values to different steps.

Does the lowest cost guarantee the best real outcome?

No. The conclusion applies to the modeled objective and assumptions, which may omit relevant factors.

AI Glossary / Graphs and symbolic AI / Beginner

Action cost function

Definition

An action cost function assigns a numerical cost to taking an action in a planning or search problem. Depending on the model, the cost can depend on the action, the starting state, and the resulting state. It tells the system how expensive a step is according to the objective being optimized.

What it means

Reaching a goal does not necessarily identify the best plan. Several action sequences may succeed, while consuming different amounts of time, money, energy, or other resources. A cost function makes the chosen comparison explicit.

The word “cost” need not mean a financial charge. If every action costs one unit, minimizing total cost means minimizing the number of actions. If costs measure minutes, the planner instead favors the smallest modeled time total.

The cost function expresses a modeling choice. A planner that finds the cheapest plan under that function has not automatically found the best business decision in every respect. Important requirements may need to appear as constraints, additional objectives, or a different cost model.

How it works

Define the states and permitted actions first. Then specify a cost for each relevant transition, often written c(s, a, s′): the cost of action a when moving from state s to state s′. Simpler models may use just c(a).

In an additive planning problem, the cost of a complete path is the sum of its action costs. A search algorithm uses those values to compare alternatives. The cost already incurred is different from a heuristic estimate of what remains before reaching the goal.

Use consistent units and include contextual information when it changes costs. If operating a machine after a different product requires cleaning, either the state or another appropriate model component must represent the previous setup. Otherwise the same action may appear to have a misleading fixed cost.

Check the search method's assumptions. Some algorithms require nonnegative costs, and some completeness guarantees need stronger conditions. Negative cycles can make an ordinary minimum-cost objective ill-defined, because repeating a cycle keeps reducing its total.

Three practical examples

1. A warehouse robot assigns estimated energy consumption to each permitted movement. The least-energy route may differ from the route containing the fewest movements.

2. A support workflow assigns expected handling minutes to available processing steps. The model distinguishes a short automated check from a longer manual review while retaining any mandatory review rules.

3. A production planner includes setup costs when switching equipment configurations. Representing the current configuration lets it compare action sequences that have identical final outputs but different preparation effort.

Think of it this way

Think of a shopping list with a price beside each item. Adding the prices gives a total under the assumption that those listed prices apply and there are no unmodeled discounts or fees.

The analogy is limited because action costs can depend on the state created by earlier actions. Buying two ordinary items is often independent; performing one operation may make a later operation cheaper, more expensive, or impossible. A planning model must represent those dependencies rather than rely on a static list when it is inadequate.

Worked example

Consider an original toy search problem with three complete routes from start S to goal G. Route A uses two actions costing 4 and 7 minutes, for a total of 11. Route B uses three actions costing 3, 2, and 3 minutes, totaling 8. Route C uses one action costing 10 minutes.

With action count as the objective, C is best because it has one step. With the stated time costs, B is best because 8 is less than 10 and 11. The goal is identical; changing the cost function changes the preferred solution.

Now suppose B requires an additional five-minute setup that was omitted. Its corrected total is 13 minutes, so C becomes the quickest of these three alternatives. This correction is to the model, not evidence that addition or search failed. The example assumes sequential actions and no overlap; elapsed time for parallel operations requires a suitable scheduling model.

Illustrative case study

Fictional case: Meadow Assembly introduces a planner that minimizes machine processing time. It repeatedly chooses a sequence requiring frequent tool changes, although those changes consume substantial operator time.

The team records tool configuration in the state and adds the applicable setup costs to the model. It also keeps mandatory inspection as a constraint, preventing the system from trading it away merely to reduce a score. This invented case illustrates objective design and missing costs, without claiming a measured productivity result.

Common misconceptions and mistakes

An action cost is not necessarily a probability, reward, or heuristic. Those quantities have different roles, even when a particular formulation relates them mathematically.

Fewer actions do not always mean lower total cost. Likewise, the cheapest next action can lead to an expensive complete plan.

Combining minutes and currency by simple addition has no clear meaning without an explicit conversion or scoring convention. Assigning a very high cost to a prohibited action also differs from making that action unavailable. Reviewers should know which requirements are absolute and which are preferences the model may trade off.

Business use

The cost function is a useful place for business and technical teams to agree on what “better” means. It exposes assumptions that can otherwise remain hidden behind a claim that an AI has optimized a process.

Before deployment, compare modeled costs with representative observed operations and examine omitted setup, waiting, rework, and coordination costs. Keep estimates separate from measured values. When costs change, review whether previously preferred plans remain sensible and whether the optimizer is using the updated model.

Check your understanding

One plan has two actions costing 6 each; another has four actions costing 2 each. Which has lower additive cost?

Answer: The four-action plan costs 8, compared with 12 for the two-action plan. Counting actions alone would give the wrong comparison for these costs.

Related terms

Heuristic
Objective function
Uniform-cost search
Path cost
search problem
planning
constraint

Learn next

Next, compare uniform-cost search with breadth-first search on a small graph with unequal edge costs. Then identify which costs depend on state and which can safely remain fixed per action.

Sources

Poole and Mackworth, Formalizing Graph Searching, including additive path costs: https://artint.info/html1e/ArtInt_50.html

Unified Planning, primary metric implementation and documentation for action-cost objectives: https://unified-planning.readthedocs.io/en/latest/_modules/unified_planning/model/metrics.html

This entry uses the planning and search sense. Numerical examples are original and make their aggregation assumptions explicit.

Previous: Action Next: Action model learning