Utility action selection

Owns per-agent scoring weights and response curves rather than world state. Each decision cycle it enumerates candidate actions, scores each from inputs such as need levels, distance, personality, risk, and context passed through response curves and combined by product or weighted sum, then selects the top action or samples among high scorers. Dill and Mark (GDC 2010) present response curves, population distributions, and weighted random selection as a fix for the abrupt switches of if-then logic. The Sims picks the behavior that most increases happiness from object advertisements; The Sims 3 (Richard Evans) samples with a modified Boltzmann distribution whose temperature rises when a Sim is doing poorly; the Infinite Axis Utility System (Mark and Lewis, GDC 2015) multiplies normalized considerations. Inputs: candidate actions and their considerations. Outputs: the chosen action and, if exposed, a ranked list of reasons.

Agents look like they are weighing options: they handle the urgent thing first but sometimes surprise. The player steers them indirectly by changing inputs such as needs, objects, and priorities, and reads odd choices as personality.

• response curve shape per consideration
• combination rule (product, sum, max) and normalization
• selection rule: argmax, random among the top N, or softmax with a temperature
• inertia or commitment bonus that prevents switching between close options
• per-agent personality weights

• Dithering: two actions with close scores alternate, so agents start and abandon tasks.
• One consideration dominates the product or sum, and agents ignore everything else.
• Choices are hard to explain because each score is the product of many curves, so players see randomness.
• Weight tuning grows with the number of actions times considerations and resists local fixes.

• Priorities emerge from scores rather than authored order, so new situations get plausible, unscripted responses.
• Coupled with needs decay, agents cycle through activities in rhythms that read as routines.
• Small weight changes ripple across a population, so balance patches visibly change agent culture.

• Can language-model agents propose candidate actions that a utility layer then scores, combining open-ended options with predictable tuning?
• Can scoring functions learned from player-labeled preferences replace hand-built response curves?