A measure of the effectiveness of the agent function. Evaluates how successful the agent’s behavior is.
Consequentialism
Evaluation based only on the consequences of an agent’s behaviors. Only outcomes are evaluated and not the actions.
Design Criteria
Performance measure design depends on the task environment. A good measure:
- Reflects what the agent is actually supposed to achieve
- Is defined over environment states, not the agent’s internal actions
- Avoids rewarding proxy behaviors that score high without achieving the goal
Task-specific Examples
- Vacuum cleaner: amount of dirt removed per unit time, not number of actions taken
- Chess agent: win/loss/draw outcome, not number of pieces captured
- Medical diagnosis agent: patient health outcomes, not number of diagnoses made
Relation to Reward and Utility
- Reward
Immediate signal given after each action. Used in reinforcement learning. - Utility
Long-term cumulative measure of satisfaction across a sequence of states. - Performance measure
External objective criterion used to evaluate agent behavior, typically defined by the designer.