Alphago was actually trained directly on game state (plus some extra computed state like "how many liberties will I have if I play this move" or "will I win this ladder"). A huge number of pro games (and countless amateur games) are available on servers like KGS in a nice computer-digestible format.