Empirical Game-Theoretic Analysis (EGTA) combines agent-based simulation with game-theoretic reasoning for analysis of complex strategic scenarios. A core step of EGTA is inducing a game model (the "empirical game") from simulation data. I survey the ways machine learning has been employed for EGTA, and present recent threads of research on game model learning that extend beyond normal-form models. Examples are approaches that discover compact game structure, learn mean-field games, or estimate extensive-form models. EGTA methods can also be extended to learn parameterized game families, and exploit game family models for empirical mechanism design.