reddit: the front page of the internet

An interesting aspect of trying to build AI's for more and more complex games such as MTG, Civilization etc. is that implementing the game logic is starting to become a considerable and even quite advanced project in an of itself. I guess this is where we face a bit of a cliff we need to jump over, if our model needs millions of games to work then we need a complex emulation system that can perfectly simulate the game with instant actions. The way to avoid this is to advance to a point where AI can'play' the game as a human and learn from the same number of examples. Not sure if any alternatives exist.