Discussion Forum

Does anyone have connections to the MIT, alpha ai project on magic ?

I hear that they are stumped by the game and that their ai is stumped too.

I somehow expect they took a look at the freaking rulebook on the web and declared the game impossible to be played by the ai, because testing if it really could beat magic demands that the MIT-team first program the rule-frames of the game, and that's years of work that they avoid by declaring it undoable.

Since I constantly create decks with evolution, and have had at least 10 years with simulations on mana and have used microproses ai as a platform to develop decks with evolution as well, I think they could need a hint or two on how to improve on their alpha algorithm if they one day decide to try it anyways.

For one thing, if the ai gains points for doing random actions, then it will loop in any deck with an infinite combo, so points for a given action should be decreased if it has been done before during a turn. That would lead to the ai getting "bored" of some actions, so it will eventually leave the loop.

In one of my own simulations the order that the deck was written in also held a manual for the order that spells should be cast in, which meant that when evolution changed anything within a deck, small alterations would happen to the casting order of spells (and lands) so any deck with black mana and lots of cheap discard would list swamps, then thought seize. If the meta adapted to play decks with kobolds to avoid discard, then the playlist would slowly evolve to play mountains first, and bolts second. A complex combo deck would also build up a playlist. In addition to this procedure some other behaviour could be set up with a list of prioritized behaviour, like only playing channel if you have a fireball.

Originally I developed these guidelines when I was considering to build a chess playing ai.
By letting evolution develop a order of moves supported by conditions the result would be more and more complex play moves.

So if you know anyone on the mit ai team that tries to crack magic, give them a helping hand and send them my way..
Posted 26 April 2020 at 23:07

Permalink

There is an AI for magic called sparky in arena, but I think that it's impossible to program an AI to be better than a human without stepping into AGI territory. We as a species probably won't see that for 20 more years if we see it at all.
0
Posted 16 April 2021 at 11:58

Permalink

I had an older project with evolution involving microproses AI.
Some decks it mastered
-1
Posted 21 June 2021 at 22:01

Permalink

In the past I've had pretty good experiences with the microprose AI. While I'm not certain it was a "real" AI, because it didn't learn from games.

Since I used evolution to generate decks even that far back, something interesting happened.

Several types of decks evolved, some were playable by both of us, while other decks specialuzed themselves at being most playable by one of us.

So sparky could become a really nasty player if it is given decks that it's better at playing than humans will be.
-1
Posted 21 June 2021 at 23:26

Permalink

wickeddarkman has deleted this comment.
Posted 21 June 2021 at 23:26

Permalink

wickeddarkman has deleted this comment.
Posted 21 June 2021 at 23:26

Permalink

wickeddarkman has deleted this comment.
Posted 21 June 2021 at 23:27

Permalink