TECH

Google's artificial intelligence wins humans in Quake III Arena

Video games continue to be used as effective but challenging tools for training artificial intelligence systems. Recently, Elon Musk's OpenAI smart bot team wins teams of players in Dota 2, an action and strategy title on the eSports route. Now Google is demonstrating the effectiveness of its DeepMind system using a first person action game (FPS), Quake III Arena.Rather than organizing a team of intelligent bots capable of organizing and defeating human players, the system had to learn and adapt in real-time to the battles. This is because the researchers' challenge was to teach AI to work in teams through randomly generated maps, that is, with each new start the structure of the elements changes, preventing agents from memorizing strategic locations.The chosen game mode was Capture the Flag, whose goal is to find a flag placed on the map and transport it to the base of the team. The rules are simple, but the dynamics are more complicated, so the bot team has been forced to learn the strategies needed to win the games as real players do. Watch the video for an AI demonstration.
According to a note from the researchers on their blog, AI agents were never instructed on the rules of the game, just the fundamental concepts of the game. As the team won the games, their behavior was reinforced. DeepMind even organized a tournament between teams of AI and AI, concluding that the "smart athletes" not only had a greater success in the victories but were considered more collaborative with each other.
Finally, as a curiosity, the researchers found that AI bots also practice the biggest "scourge" of competitive games: "camping." That is, they were quiet in a strategic place waiting for their opponents. Has Google built the first cheating AI bots?



Sapo