0
$\begingroup$

When an AI is trained to play an opposing game, such as chess or go, it can become very strong.

I have read in an article (non-scientific) the claim that AI strategies were identified by scientists while an AI was bound to play go games, as well as starcraft games. However it did not tell what these strategies actually were, how they were identified, nor did it explain the configuration in which AI played (AI vs AI? AI vs human?)

Can someone explain it to me? I am familiar with go, not with starcraft, so an explanation about go is appreciated.

I also note that the chess game is not mentioned. Is there any specific feature for chess that makes them inappropriate for strategies? Or is it the behavior of an AI in the chess game that does not allow to identify strategy?

I understand there are plenty of definitions for strategy, and the article did not give one. So let's focus on following significance: Strategy is a group of principles that tell which fields are important to fight for and which are not. A strategy gives long term rewards, which is opposite to tactics with short term rewards obtained thanks to calculation on a specific issue. With this definition, go game stand for a strategic game with a few, well known tactical situations such as line versus line.

$\endgroup$
  • 1
    $\begingroup$ Can you provide a link to the article you mention? $\endgroup$ – SpiderRico Aug 19 at 19:16
  • $\begingroup$ Welcome to SE:AI I know what your are referencing. (Players like Lee Sedol commented we can learn much from these algorithms.) A link to the article, or other articles referencing this notion would definitely be helpful though. $\endgroup$ – DukeZhou Aug 19 at 20:32
1
$\begingroup$

Truth be told, I have no idea how to play Go, but luckily this is a AI forum and not a Go forum. Addressing your questions about the specific strategies that AI discovered, there's a paper released by OpenAI titled "Mastering the game of Go without human knowledge" (https://deepmind.com/research/publications/mastering-game-go-without-human-knowledge). Here, there's a section titled "Knowledge learned by AlphaGo Zero." In it, they mention that AlphaGo Zero discovered some variants of common corner sequences. I think it may be worth checking out. The section also lists some other common Go strategies that AlphaGo Zero learned. I believe during the training process, AlphaGo Zero was trained using self-play, where the AI played itself over and over again to get better.

To address your questions about chess. I don't think there's anything limiting AI from learning strategy in chess or anything about chess that limits AI from learning an effective strategy. Check out this paper for more: https://deepmind.com/research/publications/Mastering-Atari-Go-Chess-and-Shogi-by-Planning-with-a-Learned-Model Here the authors present a general reinforcement learning model capable of mastering multiple common board games. I'm not too familiar with strategies in Star Craft, however, here's a link containing some other projects where strategies that emerge by the AI are shown: https://openai.com/blog/competitive-self-play/

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.