Hacker News Reader

Comment by shapiro92 | Hacker News Reader

but those interactions are actually not as large as a real time game such as Dota or even Go. Your deck has an X amount of cards and the synergies can be easily calculated from a machine. The real problem would be, is it smart enough to beat an unknown opponent deck?

View on HN

As a Magic player (since 1996) and Go player (since 2004) and also deep learning scientist, the main issue with Magic is modeling the game. The interactions and exceptions are really numerous, I expect Magic Arena and Magic Online are a mess of spaghetti code.

For go modeling basically you have in input an image of the go board, and for DOTA an image of your screen and the action you have are "limited" and do not change wildly depending of the state of your screen (except if you're dead on Dota).

For Magic, you can't just use deep reinforcement learning with an image input, you need to somehow track the state of your deck, cards, instant effects, what your opponent did ...

Arena at least has the advantage of not including all-cards-ever, only going a few sets back relative to release. That strips out some of the worst of them.

I suggest you look at projects like Shandalar 2012 and such - the original game is from the Banding era, but the 'AI' isn't too bad.

and the synergies can be easily calculated from a machine

This article is all about how those synergies are Turing complete and thus cannot be easily calculated since doing so reduces down to solving the halting problem.

edit: that being said I suspect that writing an AI that plays "well enough" to beat human players, as opposed to optimally, is quite doable.

Reply on news.ycombinator.com