As part of its ongoing AI research, DeepMind, Google‘s machine-learning research subsidiary, announced this week that experimental releases of AlphaStar, its StarCraft II agent, will soon be playing a small number of games. in the competitive market. For scientific testing purposes, DeepMind will analyze AlphaStar’s performance by playing anonymously in a series of blind games. This means that the StarCraft community will not know which matches AlphaStar is playing, which will ensure that all games will be played under the same conditions.
AlphaStar is Google’s DeepMind AI designed to play the StarCraft II real-time strategy game. StarCraft is a game that requires players to collect resources, build dozens of military units, and use them to try to destroy their opponents. StarCraft is particularly difficult for an AI because players must execute long-range plans over several minutes of play, polishing them on the fly against enemy counterattacks. The need to balance short- and long-term goals and adapt to unforeseen situations poses a significant challenge to often fragile and inflexible systems. Mastering this problem requires advances in several challenges of AI research, including :
– game theory: StarCraft is a game where, unlike the chifoumi (stone-paper-scissors), there is no better strategy. As such, an AI training process must continually explore and expand the boundaries of strategic knowledge.
– imperfect information: unlike games like chess or Go where players see everything, crucial information is hidden from StarCraft players and must be actively discovered by performing “recognitions”.
– Long-term planning: like many practical problems, the causes and effects are not instantaneous. Games can also take up to an hour, which means that actions taken at the beginning of the game may not be profitable for a long time.
– Real Time: Unlike traditional board games where players alternate subsequent moves, StarCraft players must perform actions continuously as the clock progresses.
– Large space of action: hundreds of different units and buildings must be controlled simultaneously, in real time, thus creating a combinatorial space of possibilities. In addition, the actions are hierarchical and can be modified and augmented.
Last January, the company announced that its AI has managed to beat professional StarCraft II players. DeepMind said that before its own efforts, no one had been able to design a StarCraft artificial intelligence that was as good as the best players. Thus, as part of his research to improve the performance of AlphaStar, DeepMind has again decided to play his AI against human players, but this time anonymously. According to the DeepMind announcement, several experimental versions of its AlphaStar IA will participate in the performance tests. They will play a number of games in Europe only.
Source : StarCraft II