Comparison between neural networks and gravity models in trip distribution : data availability vs. the performance in trip distribution. Transportation engineers are commonly faced with the question of how to extract informatie from sparse field data. For example, modelling the distribution of trips between regions, trip distribution, is highly complex and dependent on data. The aim of this study is to explore the performances of neural networks in trip distribution modelling and to compare the results with more commonly used doubly constrained gravity models. The approach differs from other research in several respects; a.o. the use of both synthetic data as well as real world data. A synthetic network combined with synthesised impedances (skim matrix) and synthesised trip attraction and production values are trip distribution modelling inputs. Well-defined differences between origin-destination data complexity increase the controllability of the test; differences in results can easily be attributed to the built up of the matrices. Data of the Rotterdam Rijnmond region is used for comparison of the model performances of real world and synthetic cases. Neural networks out-perform gravity models when data is sparse in both synthesised as well as real world cases. Sample size for statistical significant results is much lower for neural networks. In general, the results show the strong dependence of performance on the availability and quality of data. (Author/publisher)
Abstract