starcraft agents

The StarCraft AI project was the assignment in the cognitive science program where the philosophical questions about agency and decision-making met actual implementation. The task was to program an agent that could play StarCraft — or rather, a simplified version of the strategic environment — making decisions about resource collection, unit production, and combat engagement without human intervention. The gap between what a reasonable agent description sounded like and what actually producing an agent required was where the course delivered most of its content. You could describe the agent's goal state clearly. The question of how it was supposed to pursue that goal in an environment where information was incomplete and the opponent was adaptive was a different problem.

What the StarCraft context made concrete was the difference between reactive agents and deliberative agents — and the conditions under which each performed better. A purely reactive agent, following rules of the form 'if X then do Y,' could handle the fast local decisions that StarCraft required but failed at the strategic coordination that required planning across time. A purely deliberative agent, building world models and planning sequences of actions, could reason about strategy but was too slow to respond to immediate tactical threats. The architectures that worked were hybrids, and the design of the hybrid was the actual problem. That problem had a structural parallel to questions we'd been reading about in the cognition literature: how biological agents coordinate fast reactive responses with slower deliberative reasoning, and what the organizational relationship between those two systems looks like.