Games Agents Play

You are playing Civilization. Your city can build a granary or train another unit. Food will help it grow; the unit might keep it alive. The rule is easy to understand. The choice depends on everything happening around you.

Sid Meier describes gameplay as:

“a series of interesting decisions”

— Sid Meier, Interesting Decisions, GDC 2012

That is the role I want Games to play in Simulacra. A simple world can produce difficult choices: use a resource today and damage a trading relationship tomorrow; share information with a neighbour who later becomes a competitor. We can give artificial agents those choices and watch what the rules encourage.

Consider a proposed game in which agents deliver supplies between two settlements. Rewarding arrivals alone might encourage them to neglect costly routes. Rewarding speed might encourage them to abandon difficult requests. The task’s specification includes both the score and the actions available to pursue it. Specification gaming occurs when an agent obtains a good score by exploiting the rule while failing the designer’s intended task.

I would begin with understandable mechanics and record completed deliveries, unmet needs, and resource use separately from the reward. Changing the score would then be an experiment in its own right. The gap between what is rewarded and what people want would remain visible.

There is an older computational tradition behind this approach. Anand Rao’s AgentSpeak(L), introduced in 1996, represented an agent in terms of beliefs, desires, and intentions: information it holds, goals it pursues, and plans it has committed to. Systems such as Jason extended that approach to agents that communicate with one another.

Language-model agents have a less neatly specified decision process and can use a much wider range of tools. A controlled world still gives us a way to compare their behaviour when the same rule or resource constraint changes.

Games supplies those worlds. Ecology studies patterns of interaction inside them. Genus follows differences in capabilities and inheritance between their participants. Each series can examine the same run from a different angle, with conclusions limited by what the experiment actually measures.

References

Sid Meier, “Interesting Decisions,” Game Developers Conference (2012); Richard Rouse III, interview with Sid Meier in Game Design: Theory & Practice, 2nd ed.; Anand S. Rao, “AgentSpeak(L): BDI Agents Speak Out in a Logical Computable Language,” MAAMAW (1996), 42–55; Jomi F. Hübner and Rafael H. Bordini, Jason / AgentSpeak documentation.