Learning Patterns In Multi Robot Soccer Games

In this project, Professor Manuela Veloso and I work on how to learn the underlying set of rules that govern the behavior of autonomous multi-agent systems, where the role of an agent can most reliably be deduced from its physical location. The location data is collected over a continuous period of time during which we assume that the system randomly demonstrates deterministic behavior patterns from a predefined set. Earlier work on activity recognition on multi-robot domains mostly focuses on identifying the actions of a system in a given time frame. However, we investigate whether we can deduce the long-term goals of a sequence of actions that are repeated frequently.

Our main contribution is how we model the spatial and temporal behavior data of a multi-agent system as instances of geometrical trajectory curves. We have implemented this approach in the robot soccer domain, where we analyze an opponent team's game to deduce the core set of plays executed. The behavior of a robot throughout a play is represented as a continuous polygonal curve, which is in fact the trajectory it follows. Using geometrical distance metrics such as Hausdorff and Frechet, we quantify the similarity between sets of trajectories.

Currently, we have completed experiments where a team can successfully map all the plays of another team to a subset of core plays. Furthermore, we have preliminary results that show such an analysis can be done as the game progresses. This online learning of strategies of an opponent can then be used to counter the plays of an opponent in the early stages of execution. We plan to submit a conference paper based on this work in near future.

This is an example of how the behavior of the robots and the ball can be simply modeled as a 2D polygonal curve. By making the basis of our comparison on the 2D geometric distance between these trajectories, we have successfully demonstrated that we can find underlying rules of our opponents' games.

Unless otherwise stated, the content of this page is licensed under Creative Commons Attribution-ShareAlike 3.0 License