Slashdot Mirror
An AI Learned Magic: the Gathering, Now Creates Thousands of New Cards
merbs writes: Reed Milewicz, a computer science researcher, wowed a major online Magic: The Gathering forum when he posted the results of an experiment to "teach" a weak AI to auto-generate Magic cards. Milewicz had trained a deep, recurrent neural network—a kind of statistical machine learning model designed to emulate the neural networks of animal brains—to "learn" the text of every Magic card currently in existence. Then he had it generate thousands of its own.
He shared a number of the bizarre "cards" his program had come up with, replete with their properly fantastical names ("Shring the Artist," "Mided Hied Parira's Scepter") and freshly invented abilities ("fuseback"). Players devoured—and cheered—the results.
He shared a number of the bizarre "cards" his program had come up with, replete with their properly fantastical names ("Shring the Artist," "Mided Hied Parira's Scepter") and freshly invented abilities ("fuseback"). Players devoured—and cheered—the results.
What I thought from the description is that a neural network was taught how to play Magic and somehow generated new cards by trying to play with them. Think for instance of a program that tries to come up with a new chess piece by coming up with a movement pattern, playing games with that piece and trying to figure out whether it'd be useful or interesting to play with.
This on the other hand looks like something like a markov chain generator. Amusing nonsense that can give humans fun ideas.
No. It made nonsensical cards early into its learning process.
Later on it made cards like this: