Google Deepmind may be dominant at the game of Go, but it’s still struggling when it comes to card games like Magic: The Gathering and Hearthstone. Still, researchers fully intend to push it away from so-called “perfect information” games and towards games that require facing an opponent with secrets. In order to do that, they have to teach the artificial intelligence the fundamental logic behind the cards themselves.
In an attempt to see whether they’d managed this trick, Deepmind’s minders recently asked the program to create new cards. It did so, but with very limited success.
Researchers fed the advanced artificial intelligence more than 10,000 Magic: The Gathering cards, coded in Java, and about 500 from Hearthstone in python. The code represented key pieces of information contained on each card, including the mana, power, rarity and even the descriptions themselves on the MTG cards. After digesting some more open source code detailing how the games work, the Deepmind was ready to try its mind at card making.
Deepmind failed to score a passing grade on its first test for Magic, though it nabbed a low D in Hearthstone. Running the evaluation algorithm Bleu, the A.I. posted 61.4 percent accuracy for MTG and 65.5 percent success for Hearthstone. The researchers did warn that the computer succeeded in recreating the cards it had evaluated in the trainings. When it hadn’t seen the card itself before, Deepmind was totally lost.
That said, Deepmind did occasionally score 100 percent. It created the Madder Bomber, as shown above, exactly as it should have. In fact, though the test set of cards contained a Mad Bomber card, the Madder Bomber deals only 3 damage instead of 6. This subtle difference is a pretty good catch, even if it kind of scrubbed the Preparation card altogether. Hearthstone cards are more straightforward than those in Magic, possibly explaining the A.I.’s better performance, even as it received a much smaller test set.
For the researchers, the test was a success, because the computer outperformed its benchmarks. And since work on the neural networks undergirding the system will continue, it’s likely that the Deepmind will improve its production of MTG and Hearthstone cards. Unless, of course, it gets bored and takes up Yugioh. Or just sells its whole collection to a second-grader and uses the cash to buy a fresh lunchbox, like most early MTG fans.
The whole project does seem like a way for geeky Googlers to bring in some extra dough or save a little money. Hearthstone packs straight break the bank, and one day Deepmind might be capable of printing their own supply. But for now, Google gamers will have to stick to Go to impress their dates.