act Independently, but interact with others and even work together – under constantly changing circumstances: The human being has this ability mostly, it is a Foundation of our complex societies. But machines can also?
often referred to as "Artificial intelligence" (AI) called Software , learn from statistics, is in principle capable of. The study, which has now been in the magazine Science published at least. Posted you has a developer Team of deep mind, the AI division of Google group Alphabet.
learning The Software, such as a person on the Input images.
The developers had programmed a Software for the game Quake III Arena, an early Multiplayer first-person shooter from 1999. Specifically, it was about the so-called Capture-the-flag mode: In this variant of the game, Teams must capture in a randomly generated landscape, the virtual flag of the opponent and to one's own base.
The Software learns how a person on the Input images. In the game Quake III, that means that, unlike in the case of Board games like "Go" is the ambient three-dimensional and complex. But above all, you are on the "trains" – the movements – of the other only incomplete information. However, the researchers of the Software facilitated the work by using a Quake-version with a slimmed-down graphics (in which, by the way, no guns or human figures appear).
KI defeated people in the first-person shooter
"For The Win" or "FTW" Software called completed 450'000 Quake-rounds, each lasting about five minutes. It was a "Multiagent"environment with multiple enemies and teammates, which makes the complex behavior necessary. Therefore, the programme not only played against a single clone of himself. Rather, a whole Population of about 30 FTW-types, which developed differently trained.
Illustration of the Quake-flag mode, on the basis of the System "FTW" trained. Image: deep mind
In this way, FTW first learned the game mechanics and concepts such as "walls" (see Video), after which, different individual and Team strategies. Thus, the Software works not just with the Trick, to develop the best strategy based on well-known terrain, took place each batch on a newly generated map, on which the Team bases, walls, and flags always in other Places were.
After a little more than 150'000 to Play FTW surpassed the level of a strong human player. The deep mind developers tested the Software but also in practice by held a Quake tournament with 40 people. Machines and people were in different combinations - and will compete against each other in mixed Teams.
The self-learned tactics of the human strategy resembled in part.
The result is The completely different FTW-types of existing computer teams won every game against human teams, but also almost all games against a combination of human and FTW. A professional Tester-the Duo was able to fully communicate with each other, and for twelve hours trying to the tactics of the FTW-Duos set, against which they played. Nevertheless, the FTW-Some 75 per cent of games won.
The self-learned tactics, the FTW applied, similar to part of the human strategy: to focus on the defense of the own base; or Waiting for opponents who want to bring the stolen flag back. The idea of a flag carrier, chase, tested the Software extensively, to discard them again.
among the advantages of the FTW, the target accuracy in Nahtreffern (80 per cent of human 48 percent), and the reaction time. People were, however, in the case of remote hits is successful, which suggests that they capture the overall situation better.
The human reality is more complicated
The researchers put it this way: "We have demonstrated that an artificial Agent that uses only pixels and game points as an Input, can play maximum learning competitive, and in a multi-agent environment." Translated: AI can learn from scratch on Multiplayer games and win.
The most exciting aspect is the collaboration: The 30-FTW-agent had been developed individually, and therefore a different focus. Nevertheless, they managed to cooperate in various combinations with each other to develop a kind of collective intelligence. However, the five percent of the cases are relevant, where the combination of man and machine against machine team won. She suggests that this Software model to adapt their behavior successfully to people and work with them together can work.
The attempt, Software habituate a kind of collective act of intelligence, is still at the beginning.
however, as with any game: a clear goal, the pursuit of all team members. The reality of human civilization, with their diverse motives and goals is, of course, more complicated.
in Addition, the researchers note that there is still a need for optimization. So it is difficult to train an entire Software Population and to receive at the end of really different types, so a "diversity" of the solution approaches. In addition, the overall optimization of the population training is done, even after relatively short-term characteristics. The attempt Software habituate a kind of collective act of intelligence, is still at the beginning. (Editorial Tamedia)
Created: 01.06.2019, 14:23 PM