Unsupervised learning-based solution of the Close Enough Dubins Orienteering Problem

Abstract

This paper reports on the application of novel unsupervised learning-based method called the Growing Self-Organizing Array (GSOA) to data collection planning with curvature-constrained paths that is motivated by surveillance missions with aerial vehicles. The planning problem is formulated as the Close Enough Dubins Orienteering Problem which combines combinatorial optimization with continuous optimization to determine the most rewarding data collection path that does not exceed the given travel budget and satisfies the motion constraints of the vehicle. The combinatorial optimization consists of selecting a subset of the most rewarding data to be collected and the schedule of data collection. On the other hand, the continuous optimization stands to determine the most suitable waypoint locations from which selected data can be collected together with the determination of the headings at the waypoints for the used Dubins vehicle model. The existing purely combinatorial approaches need to discretize the possible waypoint locations and headings into some finite sets, and the solution is computationally very demanding because the problem size is quickly increased. On the contrary, the employed GSOA performs online sampling of the waypoints and headings during the adaptation of the growing structure that represents the requested curvature-constrained data collection path. Regarding the presented results, the proposed approach provides solutions to orienteering problems with competitive quality, but it is significantly less computationally demanding.