Exploiting Prior Information in Multi-objective Route Planning

Abstract

Planning a satisfactory route for an autonomous vehicle over a complex unstructured environment with respect to multiple objectives is a time consuming task. However, there are abundant opportunities to speed up the process by exploiting prior information, either from approximations of the current problem instance or from previously solved instances of similar problems. We examine these two approaches for a set of test instances in which the competing objectives are the time taken and likelihood of detection, using real-world data sources (Digital Terrain Elevation Data and Hyperspectral data) to estimate the objectives. Five different instances of the problem are used, and initially we compare three multi-objective optimisation evolutionary algorithms (MOEA) on these instances, without involving prior information. Using the best-performing MOEA, we then evaluate two approaches that exploit prior information; a graph-based approximation method that pre-computes a collection of generic ’coarse-grained’ routes between randomly selected zones in the terrain, and a memory-based approach that uses the solutions to previous instances. In both cases the prior information is queried to find previously solved instances (or pseudo-instances, in the graph based approach) that are similar to the instance in hand, and these are then used to seed the optimisation. We find that the memory based approach is most effective, however this is only usable when prior instances are available.