A utility-based agent is an intelligent agent that evaluates possible outcomes using a utility function and selects actions according to their expected desirability. While a goal-based agent determines whether a particular goal can be achieved, a utility-based agent can compare different outcomes and determine which one is preferable. A utility function assigns numerical values to states or outcomes to represent their relative desirability according to the agent's objectives. The agent may use these values to compare alternatives involving factors such as time, cost, safety, comfort, or resource consumption. In uncertain environments, the agent can estimate the expected utility of possible actions by considering different outcomes and their probabilities. This approach supports decisions involving trade-offs, where no single outcome is best in every respect. The quality of a utility-based agent depends on how appropriately its utility function represents the objectives and preferences of the task. A poorly designed utility function can encourage undesirable behaviour even when the agent optimizes it correctly.
Utility-Based Agents
An explanation of agents that compare possible outcomes using a utility function to select actions according to their preferences.
EU(a) = Σᵢ P(oᵢ | a) U(oᵢ)
Explanation
Variables
| Symbol | Meaning |
|---|---|
EU(a) | Expected utility of action a. |
P(oᵢ | a) | Probability of outcome oᵢ occurring when action a is selected. |
U(oᵢ) | Utility value assigned to outcome oᵢ. |
Σ | Summation over the possible outcomes of the action. |
Example or application
An autonomous delivery vehicle must choose between two routes. One route is faster but has a higher risk of delay, while another is slower but more reliable. A utility-based agent can compare the expected benefits and costs of both routes and choose the one with the higher expected utility.
Text
A utility-based agent evaluates possible outcomes according to a numerical measure called a utility function. The function represents how desirable different outcomes are for the agent. By comparing utility values, the agent can select an action that provides a more favourable result. This approach is useful when several outcomes can satisfy a goal but differ in cost, safety, time, or other preferences. Under uncertainty, expected utility can be used to compare actions with different possible outcomes and probabilities.
Key Points
A utility-based agent evaluates the desirability of possible outcomes.
A utility function assigns values to represent preferences.
Utility allows different successful outcomes to be compared.
Agents can consider trade-offs involving time, cost, safety, and resources.
Expected utility supports decisions under uncertainty.
An action may have several possible outcomes with different probabilities.
The utility function should represent the agent's actual objectives.
Optimizing a poorly designed utility function can produce undesirable behaviour.
Applications
Autonomous route selection
Resource allocation and scheduling
Decision-making under uncertainty
Robotic task selection
Financial decision-support models
Multi-objective optimization systems
Learning Outcomes
Define a utility-based agent.
Explain the purpose of a utility function.
Understand how expected utility compares uncertain outcomes.
Distinguish goal-based and utility-based decision-making.
Recognize the importance of designing appropriate utility functions.
Practice Questions
What is a utility-based agent?
What is the purpose of a utility function?
How does a utility-based agent differ from a goal-based agent?
What is expected utility?
Why can an incorrectly designed utility function cause problems?
Related Concepts
Goal-Based Agents
Rational Agents and Rationality
Decision Theory
Expected Utility
Performance Measures
Common Mistakes
Assuming that utility always represents money.
Confusing a utility value with a probability.
Ignoring uncertainty when comparing possible outcomes.
Assuming that maximizing a utility function guarantees desirable behaviour if the function is poorly designed.
Sources and verification
Sources are shown as provided in the project data. Not every source has yet been independently verified. Please report errors or missing citations.
- Russell, S. and Norvig, P. Artificial Intelligence: A Modern Approach.
- von Neumann, J. and Morgenstern, O. Theory of Games and Economic Behavior.
Suggest a correction or source · More in Artificial Intelligence & Machine Learning