M6: Abstraction and Agent Action Spaces
Real games have action spaces too large for exact computation. Outcome, state, sequence, and action abstraction compress a game while preserving strategic structure. This is essential for LLM agents, whose language-output action spaces are effectively unbounded. We also study how to measure whether an agent is actually improving: benchmarking, statistical methods, and confidence intervals. Without sound evaluation, you cannot tell whether an LLM agent improved or merely got lucky.
Module preview
Real games have action spaces too large for exact computation. Outcome, state, sequence, and action abstraction compress a game while preserving strategic structure. This is essential for LLM agents, whose language-output action spaces are effectively unbounded. We also study how to measure whether an agent is actually improving: benchmarking, statistical methods, and confidence intervals. Without sound evaluation, you cannot tell whether an LLM agent improved or merely got lucky.
Lectures and materials
L13: Agent Abstraction
Compressing strategic spaces without losing the choices that matter.
L14: Agent Evaluation
Benchmarks, uncertainty, and reliable empirical comparisons.