11 th International Congress on Advances in Civil Engineering, 21-25 October 2014 Istanbul Technical University, Istanbul, Turkey Learning to Predict: Automated Management and Correction of Prediction Techniques for Traffic Flows within a Self-organised Traffic Control System S. Tomforde 1 , M. Sommer 1 , J. Hähner 1 1 Organic Computing Group, University of Augsburg, Augsburg, Germany, [sven.tomforde|matthias.sommer|joerg.haehner]@informatik.uni-augsburg.de Abstract Individual vehicle-based traffic is still increasing world-wide. This has large environmental impact and results in serious congestion problems. Therefore, traffic-adaptive control systems and intelligent traffic management systems aim at improving the utilisation of the existing infrastructure and finding mechanisms to reduce pollutions. One successful system from this spectrum is the Organic Traffic Control (OTC) system. Fast and correct decisions are a key for efficient traffic control and management. Therefore, the decisions in OTC rely on predictions of the upcoming traffic conditions. A variety of techniques are capable of predicting traffic flows – with different strengths and weaknesses. This paper describes a novel approach to derive the best possible forecasts of traffic conditions for the near future. Thereby, a large set of prediction techniques is performed simultaneously. OTC is able to learn the currently most reliable technique and to calculate the most probable traffic condition out of the results of the different techniques. Hence, traffic engineers in OTC are unburdened from choosing the best technique and finding the best possible configuration. The benefit of the approach is demonstrated in simulations of urban areas in Germany. The simulation models reflect their actual topologies and control strategies as well as the traffic data measured in a census by local authorities. Keywords: Traffic prediction, machine learning, resilient traffic control, Organic Computing. 1 Introduction Continuously increasing traffic volumes and limited space in urban areas demand an optimisation of the existing road infrastructure’s utilisation. Thereby, the negative effects of traffic (e.g. pollution emission and fuel consumption) have to be reduced. Urban road networks are characterised by their great number of signalised intersections. The Organic Traffic Control (OTC) system (Prothmann et al., 2009) has been developed to optimise the control strategies at intersections dynamically at runtime and with respect to changing traffic conditions. Due to scalability and performance reasons, OTC works on the basis of self-organisation and has a decision horizon restricted to the controlled intersections. To achieve a network-wide optimisation, OTC further incorporates a decentralised and traffic-responsive establishment of Progressive Signal Systems (Tomforde et al., 2008) and route recommendations (Prothmann et al., 2012) reflecting the current traffic conditions within the road network. The performance of a traffic-responsive control strategy depends on the correctness of sensor data and the early detection of changes in the underlying conditions. Therefore, prediction techniques are used to estimate the expected traffic conditions for the (near) future. Based on these predictions, traffic control strategies, routing decisions and coordination patterns such as Progressive Signal Systems can be adapted in advance and consequently, their performance can be improved. Due to the possible impact of predictions, academia and economy spent large efforts on finding suitable mechanisms – which resulted in a variety of possible techniques with different strengths and weaknesses. When integrating prediction techniques in e.g. control strategies, traffic engineers face the problem to choose the best technique and to configure it as effectively as possible. This paper describes a novel approach to relieve the engineer from such complex problems related to predictions. Within the OTC system, the tasks to select the best possible prediction technique and to configure it are 1