Eg. in a well-known game it beated Stockfish (the strongest traditional chess engine), brilliantly sacrificing 3 pawns for a non-obvious, human-like compensation.
This once again challenged the notion of what the perfect play boils down to in chess.
Oversimplifying a bit, it had already been assumed that it's solely about grinding the opponent down by mercilessly accumulating miniscule advantages, while all the flashiness only worked because human players are error-prone.
As it turns out, it may have been just another limitation of chess engines as we knew (and built) them.
Also a nice game: https://www.youtube.com/watch?v=lFXJWPhDsSY
[…]
I think you could do this by changing the value function”
The essential thing with AlphaZero is that it doesn’t have a value function that you can change. It learns it from the games it plays against itself.
The best you could do is train it by playing against some algorithm that plays aggressive romantic style (playing against humans playing aggressive romantic style would work, too, but that would make for slow training)
I doubt that would produce an AlphaZero that itself plays aggressive romantic style, though.
Edit: one thing that _might_ work is to, during training, change the rules of chess to “both players lose if the game lasts over N moves” for some small N (40?). That should put incentives on aggressive play. Disadvantage would be that AlphaZero would learn very little about endgames.
This was what made the romantic style possible in the first place: humans perform worse under pressure, so they tend to crack once they're put on the ropes.
It's psychologically easier to attack than to defend accurately
AlphaZero would not have trouble finding the (likely fairly dull) winning moves, and certainly wouldn’t see any reason to play such a move itself.
I think you could do this by changing the value function:
- reward short game over long games,
- draw reward closer to loss reward,
- reward capture more.
In fact, I would like to play against AlphaImmortal specially trained to play against my Elo-rating level opponents with 51% win rate. That would be fun!