Optimizing Routes of Heterogenous Unmanned Systems using Supervised Learning in a Multi-Agent Framework: A computational study

Fast-paced but power-hungry Unmanned Aerial Vehicles (UAV) may collaborate with slow-paced Unmanned Ground Vehicles (UGV) acting as mobile recharging depots to perform large-scale and long-duration tasks such as in disaster relief management. It is important to be able to create high-quality vehicle routes in a short span of time to enable in-field, real-time deployment. A two-level optimization enables a tractable approach to solving such NP-hard combinatorial optimization problems, and it consists of an outer-level UGV route optimization to compute recharge locations and an inner-level UAV optimization to compute a sequence of nodes and recharging nodes to be visited. We consider various approaches to solve this two-level optimization. Method 1: Genetic algorithm for global search of UGV routes and a constraint programming solver for the UAV routes. Method 2: A-Teams framework that uses a combination of Genetic Algorithm (GA) for global search and Nelder-Mead for local search for UGV routes and constraint programming for inner routes, Method 3: Our proposed A-Teams by adding a supervised learning prediction step to veto out suboptimal evolutions from GA to the UGV route and constraint programming for the inner level, Method 4: Our proposed A-Teams for outer level but with a mixed integer programming solver for the inner-level. Our results on test cases show that the proposed Method 3 produces an optimal solution 30 % faster than Method 2, 79 % faster than Method 1, and 83% faster than Method 4 while being within 2 % of the solution optimality across these methods.