Google's artificial intelligence acquires "human skills" when playing Quake III

This is not the first time that we hear that an artificial intelligence has learned to play a video game. By following the lead of the IA who will participate in 'The International & # 39; from Dota 2 or the one who learned to play the Atari clbadics, we find Google Deepmind, who learned to play shooter Quake III .

The great company was very impressed by how in our society individuals work in groups to achieve a common goal . Basically, cooperation and communication between independent entities to deal with a problem. This has been likened to the effects of video games, especially multiplayer. This game mode invites you to create strategies and coordinate with others to achieve victory.

Google therefore thought that it was a good idea to choose the clbadic 1999 Quake III as a test field . Well, not exactly … they took the game and created a simplified version so that it could work its artificial intelligence. They ordered him to find a way to win in a typical Capture the Flag (CTF) game. In this way, he created several individual "agents" who had to compete in 5-minute games. Here's what it looks like:

Tactical Development and Increase "Elo"

Each agent is an identical version of artificial intelligence that competes with others . There are 30 in total, and everyone learns his own style of play.

It should be noted that agents do not have general information about the game . They must defend themselves only with the information that they get from the sight in the first person of their character.

Only the instruction to win the game is given, but not the means to do it . Agents had to learn to see, act, cooperate, move, and so on. The only thing that is done is to reward when the task is done (note a point with the enemy flag).

To make everything more difficult, the map varied randomly in each of the games . This ensures that artificial intelligence uses general strategies and not just a specific card.

Agents have played with each other more than half a million five-minute games. Apparently, the results have improved because they have acquired certain skills that are considered "human". For example, developed teammate tracking skills, champion of the enemy or self-defense .

At a certain point, they managed to participate in a mixed equipment mini-tournament of IA-IA, human and human-human AI. Incredibly, the teams composed by the machine were more successful than the people .

We see in the following table how the agents improved exponentially in the Elo notation system . They started poorly with around 500 points, and were able to outperform the average human first and even the most experienced over 1300

Certainly the ultimate goal of Google does not stand out. is not teaching Deepmind to play video games ] According to them, this method can help them address more complex global problems by studying cooperation and strategy in a multiplayer game [19659011] Google Artificial Intelligence Acquires "Human Skills" While Playing Quake III “title =” L'intelligence artificielle de Google acquiert des “compétences humaines” en jouant à Quake III “src =” https://media.metrolatam.com/2018/07/06/ctffigtagging180703r01width1500-89fc912e3c904dd5db05d3804815b450.jpg “/>