A Local Approach to Forward Model Learning: Results on the Game of Life Game