Game-playing Artificial Intelligence

The combination of reinforcement learning and convolutional networks for controlling video games through deep learning techniques achieved great success in action games that do not require a predictive strategy. In these games, the system reached human level or above. However, the system failed in comparatively simple strategy games where a manageable world has to be navigated and certain tasks have to be solved. The Asian board game Go is considered the most complex strategy game of all. The enormous number of possible positions makes it difficult for Go computers to defeat human players. In contrast, chess computers have achieved impressive performances with the brute force method, in which all possible moves are simulated for a certain number of moves. While chess computers were already able to defeat human chess players in the 1990s, the game Go was long considered unreachable for AI. In 2016, the system AlphaGo managed to defeat the then best Go player in the world. AlphaGo was trained on millions of historical games. A year later, DeepMind presented AlphaGo Zero, which developed strategies and tactics that were previously unknown through reinforcement learning and playing billions of games against itself. The system achieved a level of play that could defeat its predecessor version AlphaGo in 100 games just as often. Meanwhile, AlphaZero and AlphaStar exist as further generalizations of AlphaGo Zero, which can teach themselves any strategy game.