Evaluation of automatically generated reactive planning logic for unmanned surface vehicles

Unmanned Surface Vehicles (USVs) often need to utilize high speed reactive planning to carry out certain mission tasks. Development of a robust reactive planning logic is a challenging task. We have been exploring the use of virtual environments and machine learning to automatically synthesize a reactive planning logic to block the advancement of an intruder boat toward a valuable target. An important component of our work is to evaluate the performance of the automatically generated planning logic. We have used a virtual environment based game to compare the efficiency of an automatically discovered decision tree representing a planning logic for blocking to the behavior exhibited by the human operators. During our testing we used four volunteers to play against each other and against the computer. In human against human testing, the four players took turns playing the role of the USV and the intruder. In computer against human tests the four players played the role of the intruder while computer played the role of the USV defending a target. The efficiency of the logic was measured in terms of the time delay applied on the intruder by the USV as the USV carried out blocking maneuvers to protect a target. Our preliminary results show that a genetic programming based framework is capable of generating decision trees expressing useful reactive blocking logic.