Designer-Driven Intention Recognition in an Action-Adventure Game Using Fast Forward Bayesian Models

A method is described for quickly inferring a player’s goals from low-level inputs in an action-adventure game. Fast Forward Bayesian Models (FFBMs) use the very efficient forward algorithm normally used in the Forward-Backward algorithm, but they use transition matrices and observation functions specific to the game’s current state. In experiments with both a veteran player and novice players using a Flash action adventure game, the algorithm is faster to correctly infer whether the player is attempting to kill monsters repeatedly, explore, or return to town than a finite state machine (FSM) that must wait for less ambiguous information, yet is also more robust against momentary deviations from goal-relevant behavior. Using a constant transition matrix that does not increase the probability of out-of-state transitions on finishing a goal erases the advantage over the FSM.