Community

This thread is for discussion of this week's Feature Article, which goes live Monday morning on magicthegathering.com.

As long as it doesn't block time elemental it'll be an improvement

Seeker after Chaos wrote:

From what limited playing-a-friend's-copy experience I had, the only problem existing is that the AI will never bluff.

You said it yourself, it'll beat if it has a Giant Growth in hand. But if it doesn't? It won't.

I mean, sure, it's better than Shandalar, where it will occasionally Growth YOUR guys... but yeah. I suppose the only effective way to do that would be either 1) do it totally at random some small percentage of the time, making the AI look dumb or 2) find some way to say "if you've done it once, do it again", but that doesn't solve the problem of having it make the first bluff.

Well, if the game knows what cards exist in the format, then it could occasionally pretend that it has some other card in its colors when making a decision. I.E. 50% of the time, make decisions as though it had some random card in its hand. Most of the time, those cards won't change the optimal action. However, occasionally they will, which results in a "bluff".

Alternatively, the list of cards to pretend to have might be much shorter. For example, it could only pretend to have Instants, or there could simply be a list of cards to bluff about (Giant Growth, Cancel, Shock).

How to implement a bluff? Well, the game would just have to compare the possible game states where it pretends to have the card to the ones where it does not. Then it takes the ones it likes better (or possibly some mixed probability of the two). Of course, since it's playing against itself, it would also need some strategy other than always calling bluffs... I don't know if it has such a thing right now.

Meh, considering computers programmers have even made a computer play Go better than an intermediate human player, I highly doubt anyone is going to do a whole lot better with Magic.

However, having said that, I am rather impressed at some of the stuff they've come up with to get around the AI's shortcomings. Most likely we won't have human caliber AI for Magic for another 10-20 years. By then, we'll have enough sheer processing power to actually get stuff done in a decent time frame.

For more on what he was describing with decision trees:
http://en.wikipedia.org/wiki/Alpha-beta_search
Not quite the same, as he describes the earlier optimizations for what moves to even bother examining, but close.

Interestingly, as an experienced—but decidedly amateur—Magic player myself, I've actually learned a few tricks from my own AI! One thing it's particularly good at is damaging its opponent's creatures by direct damage spells and abilities before it declares attackers. The opponent is then heavily restricted as to what they can block with, as the blockers will die. Because in the paper game, keeping track of damage on creatures is far from easy, this had simply never occurred to me.

Uh, it didn't occur to you because it's normally a bad idea?  Duels of the Planeswalkers has a Second Main Phase, right?  Zapping the creatures early gives more information to the blocker, which is bad.  (I assume you mean "damaging non-lethally" with the "keeping track of damage on creatures" comment, as obviously lethally killing potential blockers is quite good.)

* Blocker was going to let the damage through anyway: You wasted your non-lethal DD.
* Blocker was going to block no matter what: It's a push, the blockers die in each case.
* Blocker was going to block if he thought he could save his creatures: This is the big one.  The 4/4 blocks the 2/2, takes 2 damage, then you Shock it after combat.  Dead creature.  Shocking it beforehand just means that they let 2 damage through and still have a 4/4.   (You could have Shocked their head in this case anyway.)