Cooperative Localization Using Learning-Based Constrained Optimization

Loosely coupled multi-agent estimation algorithms such as covariance intersection (CI) for track-to-track fusion, and discorrelated minimum variance (DMV) and practical estimated cross-covariance minimum variance (PECMV) for cooperative localization (CL), which account for inter-agent correlations in an implicit manner, are favored from the less frequent communication aspect. However, they can be computationally too expensive to be online implementable due to the costly optimization process involved. To reduce the computational cost while maintaining the estimation accuracy, in this paper, we report the application of Machine learning (ML) techniques to substitute the optimization processes by learning their optimal solutions to reduce the computational complexity. We focus on the CL problems and propose two data-driven approaches, namely LDMV and LPECMV to generate the solutions of two different constrained optimization procedures contained in implicit CL algorithms. For LDMV, the artificial neural network (NN) technique is used to predict the scalar solution of the problem with a single inequality constraint while in LPECMV the NN works as a matrix predictor to learn the solution of a matrix optimization containing linear matrix inequality (LMI) constraints. We discuss the design of the NNs in detail to respect the constraints. The effectiveness and the generosity of the two methods are demonstrated via CL experiments. And the experimental results show that both our approaches reduce the computational cost significantly without sacrificing the localization accuracy and the estimation consistency.