26 February 2008
Raquel Ros

The action selection problem in robotic domains is challenging, specially due to the environment properties, such as imprecision, unpredictability, uncertainty, dynamism and real-time response. We have gradually tackled most of them in the past, proposing a CBR approach in the robot soccer domain. After the promising results, we now introduce a new ingredient to the problem: the adversarial component. To deal with this last component, we include new measures in the retrieval process as well as an extension of the mechanism for selecting the case to reuse. Moreover, we make use of the ability of the system to allow explicit passes between teammates in order to avoid the opponent robots to take the ball. To validate the extended approach, we have performed an empirical evaluation, both in simulation and with real robots, in two vs. two scenarios comparing our approach with respect to an individualistic reactive one.

Institution department: 
Learning Systems