Final Activity Report Summary - FLEXIBLETRAFFIC (Development of Flexible Traffic Models that Exploit Emerging Data Collection Technologies Through the Introduction of Machine Learning Concepts)
The objectives of this research can be split in two categories:
1) The first objective is two-fold: (a) to present data-driven alternatives to the typical speed-density models, especially in the context of traffic simulation, and (b) to compare their performance in terms of estimation accuracy and computational performance.
2) The second objective is defined from the fact that traffic is volatile and can change states quickly, thus making modelling a challenging task. State identification would allow state-specific modelling of traffic, which could in turn result in more accurate simulation models, as state-specific models could then be employed, presumably with superior performance. Key methodological elements used in this part of this research include model-based clustering, variable-length Markov chains, and nearest neighbour classification.
Within the context of data-driven methods, locally weighted regression, neural networks and support vector regression have been considered as an alternative to typical speed-density relationships. These approaches provide a flexible framework for speed estimation using various explanatory variables. The results of the case study (with data from two freeways in California) indicate that the data driven methods consistently outperform the typical speed-density relationship in terms of accuracy. Loess and support vector regression show the best performance in this respect.
A new paradigm for speed estimation, appropriate for traffic simulation models, has been proposed that overcomes the restriction of the classic speed-density relationship and provides a flexible framework for speed estimation using various available data sources. The approach comprises several machine learning techniques, including k-means clustering, locally weighted regression and k-nearest neighbourhood classification. Several approaches within the proposed paradigm have been empirically tested and it was found that (for a network in Irvine, Canada) they all outperform the classic speed-density relationship.
A methodology for the identification and short-term prediction of traffic state has been presented. The methodology comprises state-of-the-art components, such as model-based clustering, variable length Markov chains and nearest neighbour classification. An application of the methodology to short-term speed prediction in a freeway network in Irvine, Canada, provides encouraging results.
Traffic state identification and prediction has many possible applications in the field of motorway traffic surveillance and control, including automated incident detection and capacity estimation.
Finally, during the course of the project, and given the accelerated progress during the first year, two additional objectives were set:
3) Investigation of more efficient solution algorithms for optimisation of state-space models without analytical representation: following the analysis of the flexible computing approaches for traffic dynamics models, the next step is the integration of these models into complete traffic estimation and prediction models. These models, however, have no analytical relationship (since they are simulation-based) and may have a large number of parameters (thousands in reasonable networks). Solution approaches for such systems are still not suitable for on-line operation of such models.
4) Flexible regression forms for macroscopic analysis of road safety data: following the success of the development of flexible regression forms for traffic dynamics models, the researchers considered whether these models could also be used for macroscopic modelling of road safety data.
Clearly, both of these objectives are large, independent problems that could not be fully tackled within the scope of this project. However, an initial investigation provided interesting directions that can be pursued further within subsequent activities.
1) The first objective is two-fold: (a) to present data-driven alternatives to the typical speed-density models, especially in the context of traffic simulation, and (b) to compare their performance in terms of estimation accuracy and computational performance.
2) The second objective is defined from the fact that traffic is volatile and can change states quickly, thus making modelling a challenging task. State identification would allow state-specific modelling of traffic, which could in turn result in more accurate simulation models, as state-specific models could then be employed, presumably with superior performance. Key methodological elements used in this part of this research include model-based clustering, variable-length Markov chains, and nearest neighbour classification.
Within the context of data-driven methods, locally weighted regression, neural networks and support vector regression have been considered as an alternative to typical speed-density relationships. These approaches provide a flexible framework for speed estimation using various explanatory variables. The results of the case study (with data from two freeways in California) indicate that the data driven methods consistently outperform the typical speed-density relationship in terms of accuracy. Loess and support vector regression show the best performance in this respect.
A new paradigm for speed estimation, appropriate for traffic simulation models, has been proposed that overcomes the restriction of the classic speed-density relationship and provides a flexible framework for speed estimation using various available data sources. The approach comprises several machine learning techniques, including k-means clustering, locally weighted regression and k-nearest neighbourhood classification. Several approaches within the proposed paradigm have been empirically tested and it was found that (for a network in Irvine, Canada) they all outperform the classic speed-density relationship.
A methodology for the identification and short-term prediction of traffic state has been presented. The methodology comprises state-of-the-art components, such as model-based clustering, variable length Markov chains and nearest neighbour classification. An application of the methodology to short-term speed prediction in a freeway network in Irvine, Canada, provides encouraging results.
Traffic state identification and prediction has many possible applications in the field of motorway traffic surveillance and control, including automated incident detection and capacity estimation.
Finally, during the course of the project, and given the accelerated progress during the first year, two additional objectives were set:
3) Investigation of more efficient solution algorithms for optimisation of state-space models without analytical representation: following the analysis of the flexible computing approaches for traffic dynamics models, the next step is the integration of these models into complete traffic estimation and prediction models. These models, however, have no analytical relationship (since they are simulation-based) and may have a large number of parameters (thousands in reasonable networks). Solution approaches for such systems are still not suitable for on-line operation of such models.
4) Flexible regression forms for macroscopic analysis of road safety data: following the success of the development of flexible regression forms for traffic dynamics models, the researchers considered whether these models could also be used for macroscopic modelling of road safety data.
Clearly, both of these objectives are large, independent problems that could not be fully tackled within the scope of this project. However, an initial investigation provided interesting directions that can be pursued further within subsequent activities.