The are over 13,000 Magic: The Gathering cards, each of which fits uniquely into an incredibly rich, decades-old world of lore, rules, tokens, and tournaments. It takes years to master the game, and even then a new set of cards comes every few months to shake things up.
That’s why it’s wonderful to see what kinds of innovation and oddity a days old artificial intelligence can come up with.
MTGSalvation user Talcos is a PhD candidate researcher in computer science at the University of Alabama at Birmingham. Talcos stuides how “deep neural networks” handle classification and problem-solving tasks. A neural network is more or less a computer program that processes information like neurons in the human brain do. A network is set up to seek out patterns in data, and reinforce the ones that fit the output the network is told to optimize. This is the process that the now famous Google “Deep Dream” network uses for image recognition (except users tell the patter seeking to go awry on purpose).
The data that Talcos fed the neural network in this case was every MTG card ever printed. Talcos wanted to see if the network could “learn” the patterns intrinsic to MTG. For example, at its most basic, the network might recognize that creature type cards have an associated power and toughness number in the lower right hand corner of the card, or that some cards have reminder text to explain abilities.
But would it learn that sorceries that produce creature tokens are typically white in color, or that bigger creatures usually cost more mana to cast? If it could, the network could in theory produce an infinite number of fresh MTG cards.
“For example, you force the network to read Shakespeare over and over and eventually it can write its own texts in the style of Shakespeare,” writes Talcos on the original MTGSalvation forum thread. “I saw that and thought: hey, why not Magic cards instead of Shakespeare?”
After feeding the network the cards’ information, it did indeed create novel MTG cards after two hours of training. It’s a small triumph considering the network has no concept of MTG’s rules, math, or even the English language. But as you may suspect, some of the initial results were either unintelligible or simply broken:
So Talcos let the network train its pattern recognition skills overnight. “I noticed that the network, now more fully trained, could generate meaningful cards,” writes Talcos. “However, it also had a knack for generating profoundly useless cards.” When Talcos returned, the network was still producing nonsense, but some of the cards were not only intelligible, but playable, even novel cards:
Though trained, the network still had problems making magic. It often failed to link the ability given to a card with the meaning of that ability, for example. And the abilities that the network created itself, like “Tromple” (an amusing mangling of “Trample”) simply existed unexplained on some cards until it was illuminated by “Adakar Vitionary”:
And of course, because MTG has so much “flavor” — each of the five colors in MTG have their own style of play, most-used mechanics, naming principles, etc. — the network almost always generated bizarre card names and indecipherable “reminder texts”. However, learn it did, and some newly created card types, like a legendary spell that could only be cast once per game, are genuinely exciting:
Since Talcos created the MTG neural network, a number of other MTGSalvation users have been helping to refine it. As Angelo M. D’Argenio writes at Escapist Magazine, others have suggested cutting down the data to just one card type at a time, such as creatures, which should help the network recognize patterns and even flavor among them.
Even though the cards generated from Talcos’ neural network are a crap shoot in terms of playability, the technique or something like it could one day become a valuable tool in a stymied designer’s arsenal, or at least a way for players to play a hilariously disorienting version of a game they’ve played for years. I know that I for one would have to “have more spells” every time a creature dies.
If you want to peruse the oddly wonderful creations yourself, the Twitter account RoboRosewater (named after head of MTG R&D Mark Rosewater) posts rendered versions of Talcos’ network once a day, and Escapist Magazine has a nice collection of its favorite cards, from absurd to over-powered, here as well as a more in-depth explanation of the network’s inner workings here.
HT: MTG Salvation