39
Lessons from developing an AI to play Melvin Zhang http://magarena.github.io/

Lessons from Developing an AI to Play Magic: The Gathering

Embed Size (px)

Citation preview

Lessons from developing

an AI to play

Melvin Zhanghttp://magarena.github.io/

https://en.wikipedia.org/wiki/File:ST Battle Chess.png

http://afflictor.com/2012/09/11/chess-programs-regularly-play-at-good-amateur-level/

http://jonathanschaeffer.blogspot.sg/2012/08/chinook-twenty-years-later.html

http://www.geekwire.com/2013/ibm-takes-watson-cloud/

http://madamepickwickartblog.com/2010/06/bite-the-hand-that-feeds-you/

http://poker.srv.ualberta.ca/

http://mathworld.wolfram.com/Go.html

https://gogameguru.com/crazy-stone-computer-go-ishida-yoshio-4-stones/

https://www.youtube.com/watch?v=2dgyHsW9LtM

http://magic.wizards.com/en/events/coverage/gpsin15/father-son-2015-06-27

https://magarena.github.io

https://magarena.github.io

Know your competition

http://cardforge.org/

http://archive.wizards.com/Magic/Magazine/Article.aspx?x=mtg/daily/feature/44

Keep an eye on new ideas

Comparison of cheating AIs

At 1s thinking time:Minimax MCTS

1 0.84

Comparison of cheating AIs

At 1s thinking time:Minimax MCTS

1 0.84

At 4s thinking time:Minimax MCTS

1 1.53

Date: Sun, 4 Mar 2012 21:04:39 +0800

From: Melvin Zhang

To: P. I. Cowling

Subject: Applying MCTS to make all AI decisions

...

Our own internal comparison with the original minimax AI confirms

that MCTS has better performance than minimax.

I hope that our preliminary findings and implementation would be

helpful to you and your team in further research in this area.

Regards,

Melvin

Iterate based on user feedback

Monte-Carlo Tree Search without cheating

Issue 92: Reported by eolebigot, Nov 12, 2011

Would it be possible for Magarena to have a *non-cheating*

Monte-Carlo Tree Search?

I tend to avoid the cheating AI, because I prefer fair games,

so an MCTS that does not cheat would be great!

https://code.google.com/p/magarena/issues/detail?id=92

Date: Thu, 30 Aug 2012 12:33:13 +0100

From: P. I. Cowling

To: Melvin Zhang

Subject: Re: Applying MCTS to make all AI decisions

...

We remain interested in MCTS for Magic:The Gathering...

see our new paper at

http://scim.brad.ac.uk/staff/pdf/picowlin/CowlingWardPowley2012.pdf

Cheers,

Peter

Comparison of honest AIs

At 1s thinking time:Minimax MCTS

1 0.88 (0.84)

Comparison of honest AIs

At 1s thinking time:Minimax MCTS

1 0.88 (0.84)

At 4s thinking time:Minimax MCTS

1 1.71 (1.53)

Know your competition

Know your competition

Keep an eye on new ideas

Know your competition

Keep an eye on new ideas

Iterate based on user feedback

https://github.com/magarena/magarena

https://magarena.github.io