Artificial intelligence company Deepmind has created a StarCraft II AI named Alphastar and is planning to have it play a number of games on the Europe competitive ladder. If players want, they can choose to have the option to play against it in 1v1 games.
Alphastar is an experiment in AI and machine learning. StarCraft II is one of the most complex games to train an AI to play. While there have been great strides in training AI on video games, and different AI have mastered single-player games like Mario and competitive gaming endeavors like Quake III Arena’s Capture the Flag mode and Dota 2, even the best AI have struggled with StarCraft II. Alphastar is different though. Deepmind uses a neural net for Alphastar that is trained through actual gameplay against real people. That’s where the players come in, and why it is so important for Alphastar to be able to play against real StarCraft II players on the competitive ladder.
If players want to play against Alphastar, all they have to do is set your preference to “opt-in” on the versus screen. Unfortunately, the likelihood of being matched with Alphastar is about as likely as being matched with any other individual player – perhaps less, as Alphastar is only playing a limited number of games to make sure that the tests remain anonymous. Thus, players won’t know if their opponent is the experimental AI or not.
The anonymity is to preserve the integrity of the tests. These tests train Alphastar’s neural net through reinforcement learning, an area of machine learning that combines exploration (trying new actions) and exploitation (acting on previous knowledge). It learns by seeing how effective the new actions are in a normal game. If players knew they were playing against Alphastar, they could consciously or unconsciously train it wrong by not giving it the experience of a normal StarCraft II game.
Perhaps Deepmind’s greatest achievement is not training an AI to play StarCraft II, but it may be training an AI to play Starcraft II like a human. Previous StarCraft AI have gotten results by taking advantage of the superhuman capabilities that a computer can utilize, but Alphastar has human limitations. Alphastar perceives the game using the same camera that players use, and it can only receive information and move units within the borders of that camera at any given time. Alphastar’s actions-per-minute are limited too, as they are set at a skilled but human level after consultations with pro players following an outcry against its superhuman clicking skills. So, if you are in Europe and want to be part of important research into Artificial Intelligence, you can opt in to play with Alphastar. However, it is best to use caution. If you lose, your MMR goes down similar to any other game.
