Given that individual cards often change the rules or provide exceptions, I am not sure a formal description can exist that is actually of any use.

Quote:Original post by Kylotan
Given that individual cards often change the rules or provide exceptions, I am not sure a formal description can exist that is actually of any use.

I don't think it's impossible: just very hard. The game Puerto Rico has buildings that provide exceptions to the rules, and yet there are a couple of computer implementations that do the right thing. This is achieved by introducing intermediate stages in the turns where cards can do their jobs. It's true that the number of buildings in Puerto Rico is a lot more manageable than the number of cards in Magic, but I think they are similar in nature.

If you have played M:tG on anything beyond the "friendly" level, you would be able to immediately recognize the true area of control: card advantage.

Every combination since the "power 9" has focused on generating card advantage in one (or more) of the main areas of play (hand, library, graveyard and on the field). If you are serious about building an AI, you could use card advantage as your base heuristic. "Does this play generate card advantage in my desired area of play?".

The problem with M:tG is that there are so many cards. You might have to limit yourself to a specific expansion cycle or ruleset, just to maintain a reasonable level of control over your finished product. Otherwise, you could spend the next 15 years tinkering with appropriate weights to every card.

Hi all, I'm the author of MTG Forge. It lets you play Magic against the computer using all of the rules. It has 714 cards and also lets you play sealed and draft games. My program can download all of the card pictures, just look in the menu when the program first starts up.

The current AI is very, very basic. (Don't fall out of your seats) Basically the computer randomly chooses a card from his hand and plays it. Technically the computer tries to play the most expensive card he can, and tries to play a creature first, then a spell (a one-shot magical effect). The AI also seems a little smarter than it is because I have worked a lot on the attacking and blocking code.

The AI for each card is hardcoded into the card code. There is a method within each card named "canPlayAI():boolean" and returns true if the computer should play that card, and false if it shouldn't. While very basic, it was good enough.

I am working on version 2 which will have some sort of "look ahead" algorithm, like min-max or alpha-beta. Right now I am trying to finish up the user interface. (I hate writing gui code, it is so messy. I would rather just program cards.)

My blog (http://mtgrares.blogspot.com) chronicles my ventures into AI programming and Magic.

Java source code is included
Windows Installer - http://www.mediafire.com/?5ggxjmwvmio
Java Jar (for all other platforms) - http://www.mediafire.com/?1ttoyo321mb

--Forge

http://en.wikipedia.org/wiki/Card_advantage

Its a short wiki, but research into that concept might prove useful. It breaks down all actions into, well, virtual card advantage. Card advantage is good, and its a scaler value, so making a computer want it should be straight-forward. Tiebreaking between different actions might get tricky, but thats another ballgame altogether

#1 A DSL(domain specific language) for defining card classes + an algorithm for classifying cards. The DSL will have to be built on top of the representation of the cards themselves though.
#2 A few dozen utility ratings based on abstract positions

Considering the vast number of cards and rules, I'd say go with utility by ranking utility based on an artificial personality/playtype and then performing the generated actions which provide the most utility. As a human player, I don't know what all the cards do but I can still play any deck because I know the rules so I can predict cause-and-effect and know how to use the cards in a general sense. Special combos could be hard-coded.

It's really hard to know what an opponent will do far into the future but several thousand possible states could be generated in very little time. Using card classes would make coming up with combos on the fly(during the game) much easier. The game could learn to predict stuff by remembering opponent actions in previous turns and by using Markov models, both with specific cards and by analyzing the class of cards used.

Utility/Abstract areas of interest: number of cards in hand, total creature defense/offense, elimination of things which have damaged me/creatures in the past, mana production, having or eliminating special cards, damage to the opponent, life to self. Aspects of the game which a human player must have an appreciation and awareness of must be encoded to the ai as having utility.

I don't know if it will make a better ai player, but the card classes could also be fuzzy(in the sense of fuzzy logic).

Remember, the AI doesn't just have to play a deck (though that is certainly possible). It should also be able to build decks. Let's consider the archtype decks, and how one might weight the cards in a given situation:

Aggro: Here, the goal is to burn your opponent down before they can establish defense. Card advantage is generated on the playing area, normally by sacrificing advantage in the hand. Strong cards will cost 3 or less mana, with easy color requirements. Prefered creatures will have special abilities, such as first strike, shadow, flying, to make them unblockable. Spells tend to focus on removal of threats threat removal, rather than general advantage.

Control: The infamous "no" decks. These decks establish card advantage by simply denying advantage to the opponent. "Counterspell", "Wrath of God" and "Null Rod" are all good examples of control cards. Strong cards will cover a broad range of mana costs, covering the beginning, middle and end game. Prefered reatures and spells are weighted less for what they can do, and more what they can prevent the opponent from doing.

Combo: These decks are designed around the interaction of 2 or 3 specific cards. Card advantage is developed through a single, overwhelming move in an 'all or nothing' kind of play. Often the combo deck can be completely stopped with a well-timed counterspell. Obviously, combo cards ("stroke of genious", "oath of druids", "squee, goblin nabob") are highly valued. Also highly valued are search cards, and cards which will buy the player time.

Each card could be given a weighting based on these 3 archtypes, which would also be valuable in determining play choice.

You can now begin to build personality. Obviously, a control player will assign values very differently than an aggro or combo player, and vice-versa. Here's an example:

Let's say our AI likes to play aggro decks, and prefers black as a primary color, and prefers card advantage on the field. From these choices, the AI can look at the card color, it's weighting as an aggro card, then include or exclude it in her main deck based on whatever other criteria you might assign. Using this model, the AI will naturally be drawn to cards like "Black Knight", "Phyrexian Negator" and whatever else is hot in "black weenie" style decks.

Playstyle will evolve organically from the AI's personality. The higher score the card holds for the given playstyle, the more likely it will be the AI will try to play it under normal circumstances. Continuing with the "black weenie" example above, the AI can evaluate it's cards by which will give it the largest advantage in it's preferred area of control.

For example, "Black Knight" verses "Diabolic Edict": Both have the same mana cost, but one is a permanent, while the other is an instant. Which to play? If there are no creatures on the field, the choice is obvious.

But what if the opponent has 1 creature? In this instance, it is almost always better to play the creature first. First of all, there's summoning sickness. Aggro decks work on building early tempo, and delaying the creature's appearance delays tempo. Second, the instant can be played on either turn. It can be used defensively to remove your opponent's attack on his turn, or it can be used offensively by removing a blocker on your next turn. Third, a creature generates card advantage over time. The longer it is in play, the greater the advantage until tempo is lost. However, the instant generates no card advantage. It generates card parity - one of mine for one of his. And that parity is lost as soon as the opponent plays another creature.

And if the opponent has more than 1 creature? In this instance, We have to determine whether the card parity of casting an instant will outweigh the long-term advantage of playing a creature. Are one or more of my opponent's creatures "special purpose"? Are they costing me tempo? Am I about to die?

Given this analysis, your cards should at least include the following (hidden) values:

AggroWeight
ControlWeight
ComboWeight
AreaOfPrimaryControl (hand,library,field,graveyard)
PrimaryValue
AreaOfSecondaryControl (hand,library,field,graveyard)
SecondaryValue

Of course, this is only a starting point. There are obviously many considerations to make, even in a simple creature rush. (Is my opponent running with mass removal? Are there special conditions in play, like a "Platinum Angel"? Is this creature meant for something other than combat? Have I lost tempo? etc.)