Utility AI: make action scores comparable and decisions stable

Give each score a design meaning

Utility AI evaluates candidate actions using numerical preferences. A response curve translates a world quantity, such as fatigue or distance, into a score meaningful for a particular decision. A value between zero and one is not automatically a probability, nor does normalization make unrelated actions comparable. Graham explains the scoring approach; Lewis shows how individual considerations shape game behavior. [1][2]

Separate permission from preference

Reject unavailable actions before comparing desirable ones. A repair action that requires a missing tool should not win because urgency is high. Define whether a score combination is compensatory: a weighted sum allows a strong factor to offset a weak one, while multiplication makes a zero factor decisive. Multiplication also changes scale as more factors are added. The aggregation rule is a design choice that must be tested, not hidden inside a generic scoring helper.

Worked example: a small lead need not trigger a switch

Consider an illustrative maintenance robot. Repair urgency is 0.8 and convenience is 0.4. Its repair score is 0.6 * 0.8 + 0.4 * 0.4 = 0.64. Rest scores 0.70.

The robot is already repairing. With a switching margin of 0.08, a challenger must exceed 0.72, so the robot continues. If rest later reaches 0.80, it switches at the next permitted interruption point. If repair becomes invalid, the margin must not keep it running. These invented scores illustrate a policy; they do not establish an optimal robot personality or measured player preference.

Inspect the decision surface

Build a small decision inspector showing raw inputs, curve outputs, eligibility, combined scores and the reason for a switch. Sweep one input while holding the others fixed. Then inspect combinations: high fatigue with high urgency, distant work with nearby rest, or no available actions. Decide what the fallback does when every candidate is rejected.

Plot thresholds alongside noisy input traces. Verify that hysteresis reduces pointless switching without trapping an agent in a poor action. Specify emergency interruptions separately from ordinary preference changes. Record how long an action has run and whether it can currently stop; a selection policy cannot make an unsafe interruption valid. Keep captured situations so later curve edits can be compared against the same inputs.

Bound evaluation cost and model complexity

Score only the candidates whose information is available at the decision point, and measure expensive spatial queries separately from curve arithmetic. Cache with explicit freshness rules. Nearly identical considerations can count the same motivation repeatedly, and differently scaled action sets can introduce accidental bias. Prefer a small explainable model before adding more curves; an opaque score is difficult to tune even when the arithmetic is correct.

Sources

• [1] An Introduction to Utility Theory — David “Rez” Graham, Game AI Pro.
• [2] Choosing Effective Utility-Based Considerations — Mike Lewis, Game AI Pro 3.