Fresh off the back of beating Lee Sedol, the best Go player in the world, DeepMind have turned their attention to Hearthstone. For now though, they will not be looking to actually play the game.
Google are using Hearthstone and Magic: The Gathering cards as an exercise in language. The artificial intelligence has many variables that it can use to express what it thinks a card does. The program is given a small training set of cards, from which it tries to predict what it thinks other cards do. The published by DeepMind explains that nearly all errors come in the effects of the cards, and that the identification of card name, type, rarity, and character class are unsurprisingly very accurate.In the image below, output in green, is output that is correct. Output in red is incorrect. In the instance of , the entire card is correctly understood. DeepMind explain that this is probably because , which is very similar, was in the training set. On the other hand has very little in common with other cards. This leads to the AI guessing that it sets a target minion's attack to 3.
If Google are looking at card games from a language point of view, it strikes me that it's only a matter of time before they decide to try to play them too. Thankfully there will always be a place for human vs human world championships!