Developers of the Google AI team, Deepmind, have developed software that simulated individuals can work in teams. Artificial Intelligence For Victory (FTW) plays the arena shooter Quake 3 Arena in Capture Flag Mode, in which a team must wear the flag of the opponent at its own base. The system is waiting to work together with humans as well as other machines, to be able to adapt to the ways of playing and winning surely.
Labor Market
Robert Bosch GmbH, Hamburg
Radio 93.3 Mhz Munich Munich, Munich
FTW must has played about 450,000 simulated Quake games and learned from scratch the mechanics of the game and the work of # 39; team. The multi-instance cooperation process is what developers call leraning multi-agents. However, the developers have changed the shooter a bit. The graphics are even simpler, the figures are balls instead of real patterns and the textures are more clearly distinguished from each other. In addition, playing cards can be generated randomly. This is what the AI should adapt to any scenario and any map size as it learns.
The trained model was used in a Quake tournament with 40 players. There, the system played with and against human players. A team of machine-only players has won every game against humans. The victory rate against a team of humans and machines was 95%.
The developers were able to discover AI tactics, which also used human players – following a teammate, camping in the opposing base, defending their own base and Distraction Maneuvers while a separate player carry the flag. Interestingly enough, human players found that the AI's willingness to collaborate with fellow humans was better. The system does not know solitaire and does not verbally communicate slowly, but directly badyzes the movements of players in real time
Instagib Quake heavily modified
However, the test match seems very different from the current Quake 3 Arena . It's more like a game with an activated Instagib mutator, where players can only use a weapon that stops enemies with a hit. The whole concept of aiming and meeting is largely ignored here. Instead, AI agents usually run straight to the target – a tactic that almost always works well in Capture the Flag.
Nevertheless, the results are interesting for the future gameplay, like the creation of bots to face the players. Learning NPC opponents could create a whole new level of immersion for games.
