The problem of developing virtual opponents that are both challenging and fun to play with is not easy to tackle. It is generally accepted that one of the primary goals of virtual characters is "suspension of disbelief". It means that if a character looks like a human, it should behave like a human to keep players immersed into a game world. We experiment with human-like characters in a variety of games, mostly sports due to their simplicity. In our experiments with tennis and fighting games we could create human-like agents for "one-vs-one" scenarios. In this project, we strive to create teams of human-like agents, able to pick up play styles of their human "coaches" and substitute them in the game.
Many real-life soccer matches are recorded with systems such as Tracab and converted into digital datasets. These datasets contain information about individual players' and ball locations on the field at any given moment of time. In practice it means that they can be used to reconstruct the whole game and analyze team strategies:
We aim to adapt our AC SDK-based soccer teams (designed for an older system "TSG Simulator") for Google Research Football environment and train them on real-life player tracking data.
Project repository: SoccerGame.