beating KataGo by Adversarial attack

An abstract strategy board game for two players
Post Reply
Hiroshi Yamashita
Posts: 15
Joined: Sat Oct 30, 2021 1:31 am

beating KataGo by Adversarial attack

Post by Hiroshi Yamashita » Sat Nov 05, 2022 7:03 am

These two are recent papers about beating KataGo by anti strategy.

Adversarial Policies Beat Professional-Level Go AIs
https://arxiv.org/abs/2211.00241
Are AlphaZero-like Agents Robust to Adversarial Perturbations?
https://openreview.net/forum?id=yZ_JlZaOCzv

Rémi Coulom
Posts: 168
Joined: Tue Feb 12, 2008 8:31 pm
Contact:

Re: beating KataGo by Adversarial attack

Post by Rémi Coulom » Sat Nov 05, 2022 1:03 pm

Thanks, these are interesting results.

I have always felt that it is a fundamental weakness of the AlphaZero approach that the training set is made of self-play games: it does not learn to handle positions that do not occur between very strong players. That may be why Leela Zero had a lot of trouble learning ladders. If there is no ladder being played out in the training set, then the network cannot learn how to evaluate them.

Hiroshi Yamashita
Posts: 15
Joined: Sat Oct 30, 2021 1:31 am

Re: beating KataGo by Adversarial attack

Post by Hiroshi Yamashita » Sun Nov 06, 2022 5:19 am

I think the reason of ladder fail in LeelaZero, it needs search and long distance recognition.
In Shogi, AobaZero can not play simple capture move from to 19 to 91 by bishop.
With bishop movable position information input for NN, it can play.
But AobaZero still can not play bishop discover check by moving rook from 46 to 86, a bit more difficult move.
dlshogi(current strongest Shogi program by DCNN) also can not find this.
AlphaBeta program (Yanauraou) can find it easily.
Maybe this is limit of DCNN, and something hybrid method DCNN + AlphaBeta will be better.

Post Reply