An AI-Augmented Kalman Filter Approach to Monitoring Network Traffic Matrix

Many network operations depend on traffic matrix (TM). However, how to obtain TM often turns out as a dilemma, i.e., direct measurements have a high accuracy but with a high operation cost, while indirect measurements save measurement budgets but at the cost of lowering accuracy. In this paper, we propose an AI-Augmented Kalman filter, attempting to exploit the best of both. By viewing TM as the unknown state and indirect measurements as state observations, TM estimation is formulated as a filtering problem. Unlike the conventional Kalman filter that assumes explicit model knowledge for both state prediction and update, we learn them from direct measurements of TM and construct the AI-Augmented Kalman filter in a data-driven way. Specifically, a novel recurrent neural network called ConvGRU, is employed to capture characteristics like spatio-temporal correlations and traffic dynamics, in regard to establishing relationships for both state transition and difference reasoning of state observation. Evaluations on real datasets demonstrate that TM estimated by our proposed scheme gains superior accuracy compared with existing methods. Since both direct and indirect measurements are employed for TM monitoring, our proposed scheme also enables one to further develop a balance between performance and cost according to their practical deployments.