Automatic Generation of Evaluation Courses for Car Navigation Systems Via Combinatorial Optimization

To accelerate development and improve the quality of car navigation systems, we have built a system for automatic generation of evaluation courses. In general, the operation of car navigation systems is verified by driving tests using vehicles. The evaluation courses need to be designed so that inspection sites, such as underground parking lots, tunnels, etc., will be visited during the evaluation period. They should be circuits that include as many inspection sites as possible within a defined distance. However, as the number of the inspection sites increases, the number of courses that can be designed becomes enormous. This makes it difficult to create courses that meet all of the requirements. Hence engineers have spent a lot of time on evaluation course design. For this reason, automatic course generation has become essential for reducing man-hours. We believe that one of the effective approaches is to treat automatic evaluation course generation as a combinatorial optimization problem. In our formulation, inspection sites are grouped into clusters according to the required number of courses, and the shortest circuit is constructed in each cluster.Then, we treat the clustering and shortest circuit generation problems separately as a bi-level combinatorial optimization problem. In other words, the original problem is divided into smaller parts of the combinatorial optimization problems. We then propose a Markov chain Monte Carlo method for solving the bi-level optimization problem, and construct a system for automatic generation of evaluation courses. The proposed method significantly reduces course-design time compared to manual course construction.