Cooperative Regression-based Forecasting in Distributed Traffic Networks

The paper addresses intelligent data processing and mining in distributed multi-agent networks. We consider a distributed traffic network represented as a multi-agent system, where each agent-vehicle autonomously forecasts its travelling time. We propose a structure for such agents and their corresponding distributed data processing and mining modules. We use distributed linear and kernelbased regression models for travelling time forecasting on the basis of available values of affecting factors. To improve the forecasting quality, the agents exchange information (model parameters or observations) with other traffic participants. We describe forecasting methods, paying special attention to their implementation for streaming data. We propose a cooperative learning algorithm based on the construction of confidence limits for estimates. Simple examples of each method illustrate their application principles. A case study using real-world data from Hanover (Germany) illustrates a practical application of these methods. We propose a structure for linear and kernel-based regression models, which we implement for different multi-agent system architectures (centralised, coordinated/uncoordinated). We analyse the efficiency of linear, kernel-based and aggregated regression estimates in different architectures on the basis of their relative forecasting errors and a goodness-of-fit measure. The results demonstrate that a suitable coordination schema in a distributed architecture provides almost the same accuracy as a centralised architecture.