It is possible to use case-based reasoning for forward simulation in Mario AI, this is explained by Game engine learning from video They are using features:
distance, velocity and position
to predict following game-frames like in a physics engine. What my problem with the paper is, that the “learned game engine” seems to be working autonomously. That means, the human operator is out of the loop, he is not changing the cases manually and he is not involved in predicting the future. Is it possible to make an interactive forward simulation? For example, the game-engine asks the operator what will happen if Mario jumps over the wall.
What i do not understand is how to design the GUI-interface in such a situation, because the number of possible reactions of the game-engine to a situation is endless, and to make the simulation interactive the user would click very fast on some buttons. For example, if the desired framerate is 30fps, should the user define all the parameters in a timestep of 1/30 seconds?