Must the performance element be a neural network?

No. It can use rules, planning, a learned policy or other decision mechanisms.

Can it contain a world model?

Yes. An action selector may use a model to reason about consequences.

Does a system without online learning have one?

Yes. The acting component can use knowledge learned earlier or supplied by designers.

AI Glossary / Graphs and symbolic AI / Beginner

Performance element

Definition

The performance element is the part of a learning agent that selects actions in response to its available observations and internal state. It carries out the agent's current behavior, while other architectural components can evaluate and improve how that behavior is produced.

What it means

In the classical learning-agent architecture, the performance element answers the practical question of what to do next. It may use rules, a policy, search, a model or a combination of mechanisms. The word performance refers to performing the task, not merely measuring speed or accuracy.

The learning element modifies aspects of the agent using experience. A critic supplies evaluative feedback according to a performance standard, and a problem generator can suggest exploratory actions that produce useful experience.

These are conceptual roles. They need not correspond to four separate software services, and a modern system may distribute the roles across several components. The distinction helps explain which mechanism acts and which mechanisms change or assess that behavior.

How it works

The agent receives a percept through its observation interface. The performance element combines that input with relevant internal state and the current decision mechanism. It selects an action, which the agent's actuators or software interfaces carry out.

Experience from the interaction can then inform learning. The critic evaluates aspects of behavior or outcomes, and the learning element decides how to adjust the agent. Those adjustments may affect a policy, world model, value estimates, representations or other knowledge used by the performance element.

An exploratory suggestion can influence action selection when the design permits it. However, exploration does not erase the need for action constraints or a meaningful objective.

At a given moment, the performance element may operate using a frozen learned model. The system still has a performance element even if learning happens offline between deployments rather than continuously during each interaction.

Three practical examples

1. Recommendation: The serving system uses its current model and decision rules to select items to display. A separate training process can later update the model based on appropriate feedback.

2. Mobile robotics: A controller selects the next movement from observations and its internal map. Learning may improve the map or decision policy, but the action-selection role remains identifiable.

3. Support routing: An agent chooses a destination queue or requests human review. Evaluation of later resolution outcomes can influence subsequent policy updates without being the action selector itself.

Think of it this way

Think of the performance element as the player currently making moves, while coaching and review processes help the player improve. The player can use learned skills without receiving a coaching update on every turn.

The analogy does not imply a human-like mind or a fixed organizational separation. Software can combine acting, feedback and learning in one program while still performing these distinct functions.

Worked example

Consider a toy routing agent with two actions: send a request to Queue A or Queue B. Its performance element chooses the action with the larger estimated value for the current request type.

Initially the stored estimates are Q(A) = 0.60 and Q(B) = 0.55, so the greedy selector chooses A. Suppose an observed outcome is encoded as feedback value 0, and a simple learning rule updates only the selected estimate using step size 0.5.

The new estimate is Q(A) = 0.60 + 0.5 × (0 − 0.60) = 0.30. Q(B) remains 0.55. On the next otherwise identical decision, the same performance-element selection rule chooses B.

Action selection and learning are different steps: the selector compares the current estimates, while the learner changes one of them. The calculation illustrates the separation of roles, not a recommended production routing algorithm or proof that one outcome justifies such a large update.

Illustrative case study

In a fictional project, Willow Support describes its routing service as self-improving, but staff cannot identify which code actually changes its decisions. Investigation shows that the dashboard computes outcome statistics while the live selector continues loading an older model.

The team records the model version used for each action, documents the update process and verifies that an approved update reaches the serving component. It separately reviews whether the feedback is appropriate for the routing objective.

The case illustrates that collecting a performance score is not itself learning, and producing a new model is not the same as deploying it into the element that selects actions.

Common misconceptions and mistakes

The performance element is the performance measure. One acts; the other defines or evaluates success.
Every component that processes data is the performance element. The defining role is selecting behavior in the agent architecture.
A learning agent must update after every action. Learning can occur periodically or offline.
The critic necessarily gives the correct action. It supplies evaluative feedback; its exact form depends on the system.
Improved training results prove improved deployed behavior. The updated mechanism must reach the acting system and be evaluated in context.

Business use

The distinction is useful when reviewing an AI system's operational design. Identify which component chooses actions, what information it receives and which versions of learned artifacts it uses.

Then identify how feedback reaches the learning process and how changes reach the action selector. This exposes gaps between data collection, training, evaluation and deployment. It also clarifies where action permissions, fallback rules and review routes are enforced.

Keep the behavioral objective explicit. A performance element can consistently follow a poor policy, and a learning element can optimize feedback that does not reflect business success. Architectural clarity makes those questions easier to investigate but does not automatically answer them.

Check your understanding

Which part changed Q(A) from 0.60 to 0.30 in the example, and which part selected B afterward?

Answer: The learning step changed the estimate. The performance element applied its existing selection rule to the new estimates and selected B.

Related terms

Learn next

Draw an agent's path from observation to action, then add the feedback and update paths separately. Label which component acts, which evaluates and which changes the decision mechanism. Check whether model versions can be traced across these paths.

Sources

Russell and Norvig's learning-agent diagram distinguishes the performance element, learning element, critic and problem generator: https://aima.cs.berkeley.edu/slides-pdf/chapter02.pdf

The routing update and Willow Support case are original illustrations of architectural roles. They do not prescribe a particular learning algorithm or deployment process.

Previous: Perfect rationality Next: Performance measure