I thought the same. Ender's Game and the cell classification task have something in common: the end status to achieve (whether to eliminate all opponents or identify something that's known). If we think this way, these two tasks could still be treated as machine learnable tasks (albeit poorer solution possibly). A genetic algorithm for example, could be trained to derive the means to achieve the goals. In cases where we don't know the end goal of the game, how do we train ourselves towards an unknown objective or without guidance of known answers.