• Contact Person: Viliam Lisý
• Key Partners / Sponsorship: US Air Force
• Years: 2007 -

U.S. Air Force Research Lab Project Deep Adversary (DeepA) investigates planning and adversarial reasoning in realistic games involving multiple players of different types, strength and goals. As a part of the project, a sound formal model of adversarial behavior is being developed. The model forms a basis for advanced techniques for modelling, detecting and predicting adversarial behavior.

In the project, we focus on three basic applications of the opponent model: execution, utilization, and generation. The first application is crucial for the other ones. We develop intelligent agents that can be parameterized by the opponent model and play based on the easily adjustable characteristics.

When playing against these players, we investigate how knowing the opponent model the other players execute can help to improve performance of the agents in the game. We utilize the information about the opponent model in the planning process of the agents as well as in prediction the opponents’ actions.

In order to complete the framework, we are using machine learning and regression techniques to generate models of the opponents based on the observations made in past games or earlier in the history of the game, where we want to utilize the model.

The techniques are being evaluated using a sophisticated A-globe-based game testbed, which allows simulating complex games in realistic environments.

Some of the results of this projects were developed in collaboration with CMU in the project of the Czech Ministry of Education, Youth and Sports under grant ME09053.