Application of Proper Orthogonal Decomposition and Artificial Neural Networks to Multiple UAV Task Assignment

Abstract

In this paper, we describe a new approach to the multiple UAV task assignment problem. This approach, which we call PODANN, utilizes a corpus of optimally solved, randomized instances of a given task assignment problem. Each solved instance can be described by a combination of two vectors: a \distance vector that encodes distances between UAVs and targets and a \solution vector that encodes the optimal tour. Proper orthogonal decomposition (POD) is applied to a training corpus of solved instances in order to flnd a linear mapping to a lower dimensional space. Then, an artiflcial neural network (ANN) is trained to map each distance vector in the training corpus to the corresponding instance representation in the lower dimensional space. The solution vector can then be estimated by taking the pseudoinverse of the output of the ANN. A separate validation corpus is used to evaluate the efiectiveness of the technique. Results and shortcomings are discussed.