Just thinking of Counter Strike as a canonical example, if you have a coordinate representation of the map and for every player you have a complete history of everything you might need for every moment of time, say every 10 miliseconds or something and you just recorded where the players went around the map, and what they did. So maybe, for example, at each time t you had v(t) a vector function whose components are (location(t), velocity(t), rotationVector(t), rotationVelocity(t), isFiring(t), ammoCount(t), health(t), isEnemyVisible(t), isTeamMateVisible(t)) or whatever other measurements you think would be useful for the state space.
If you collected a bunch of data like this what sort of hypotheses or models would you consider testing? Would it be possible to construct a probabilistic model to classify histories and possibly predict outcomes of player-enemy encounters? Could you train AI bots on this data?