KataGo's new in-progress run includes a small amount of self-play training where a high playout version fights against a low-playout version while at a disadvantage (either handicap stones, or a significant reverse komi). The neural net is told whether it is the high playout side or the low playout side. So the neural net learns for the high playout side that play more aggressively works well, and for the low playout side to play more solid and secure works well, and both sides are counteroptimizing against the other - the high playout side tries to learn strategies for making the game messy that are effective even while the low playout side actively learns how to minimize the possible trouble.

The idea is that you can now use the high-playout side in handicap games for hopefully more effective handicap game play.

It seems this has resulted in a very different high-handicap style than KataGo's official released run or that of any other "zero" bots. For one thing, it no longer begins the handicap game by invading all four 3-3 points, despite continuing to like 3-3 in even games. Perhaps it's learned that 3-3 makes the game "too easy" for a weaker opponent and settles things too fast?

The new version has been up on OGS for a few days and it's been fun to watch.

Example game (Note: OGS dan ranks are quite on the soft side):



Another example game:

This is very nice! The games look like real human handicap games.

Thank you!

I've used your preceding test bot on OGS with various handicaps and board sizes, had a lot of fun with it and found it very useful to spar with. Thanks again for that! ^_^
I have a question though.
So far, I've reviewed my games with the official version of KataGo on my own computer, as I thought the expected variations from your test bot available at the end of the game wouldn't be as useful since it assumed Black to be weaker, hence to play somewhat inefficient moves...
But now, given your above explanations, I wonder if I should have not in fact taken your test bot expected variations very seriously instead, since Black would have tried to simplify the game while still playing top level moves (?)
So that's my question: how useful do you think White's expected variations are to review these handicap games?

_________________
Bored with skillful play? Watch my blunders on my YouTube Channel

I don't know! If you could try for yourself and see as a player how useful you think they are in practice, that would be a helpful thing to share.

During training, the weaker side never deliberately tries to play bad moves. It's weaker because it gets less playouts, but it is always trained to still do its best. And outside of training there is no actual reduction in playouts. The only difference is that the neural net is told "you are the stronger side" on White's moves and "you are the weaker side "on Black's moves during the search, so it gives a policy, score, and winrate predictions accordingly. You search exactly as usual and the only difference is the numbers that are coming out of the neural net.

So I would not expect a higher rate of "mistakes" than usual. The main difference is the *goal*. Normally both sides try to find moves that can achieve a position such that after that if the rest of the game were played by equal players, the predicted winrate and score will be favorable.

Instead now, white's moves will try to achieve a position such that if the remainder of the game after that position were played by a strong player against a weak player, the predicted winrate and score will be favorable for the strong player, and black's moves will try to get to a position that the predicted winrate and score will be favorable for the weak player.

So my guess is that if you learn from reported black moves, you will likely learn moves that KataGo believes are reasonable ways to try to win if you have a huge lead but you think the opponent can outread and outfight you - such as maybe solid plays and moves that keep pace without picking unnecessary fights, including perhaps moves that give up small amounts of edge that would necessitate sharper play to keep. If you do your own review, you will instead get moves KataGo believes are most likely to be good if the opponent is equally strong. Which could still bias between safe vs aggressive moves depending if you are ahead vs behind - KataGo is partly but not completely score maximizing. And maybe one could try setting it to "aggressive mode" to review handicap games you played as white, once next release is out?

I'm not sure about all of the above - I wrote the code, but it's only been a few days and it's not like I've played a lot with it!
And of course, regarding "weak player" in all the above, KataGo with low visits should still be "high dan++".

The first game is amazing, Black is stronger than me, and doesn't make any horrible blunders; yet slowly loses points by being inefficient. The next KataGo run, if it has a server implementation and continues into higher block/filter sizes could be giving pros much higher handicaps than LZ, even if it's on parity with the LZ run. Maybe even beating top pros at 4 stones. We've seen 3 stones to be difficult for current AIs, but how much of that is playing in a style that doesn't fit well with handicap games?

Thanks for the reply
After reflexion, I guess it would make more sense for me to stick with a normal version of Katago when reviewing handicap games, since that would be the only way to have a "fair" assessment of the score (i.e. not one biased towards White). Plus, as you said, normal Katago isn't completely trying to maximize the score anyway, so it should also favor safe enough variations for Black to play... as long as White didn't catch up with him, that is :p

_________________
Bored with skillful play? Watch my blunders on my YouTube Channel

Actually black's 106 and 108 are both blunders in the first, 7 handicap game.

But that doesn't take away from the fact that this KataGo is indeed playing very nice handicap style as white, the idea of self-play between low-playout vs high-playout worked amazingly well!!

Is the network available for download?

_________________
Sorin - 361points.com

106 is just "not good" and not immediately obvious why it's bad and 108 is more of the "blunder" since it tactically loses

Just curious: does this mean there's a new input to the neural net, and hence a minor change to the network architecture? Or is it "told" by some other mechanism?

Yep, new input to the neural net. The neural net readily notices the correlation between this new input and the fact that one side sometimes is winning more than they should given different positions, and then notices that the difference is even more extreme if the position is "messy" than if it's "simple", so then the policy learns to play more aggressively/defensively depending on which side it's on, and it all goes from there.

Since it requires this specific training, of course, it only works with the newer neural nets from the latest run.