Drive Time Average and Variation EstimatorProject OverviewData CollectionModel EvaluationNeural Network Models TriedModel Details:ResultsPractical InsightPractical Insight?ConclusionsDrive Time Average and Variation EstimatorBy Matthew A. HalsmerUniversity of WisconsinECE 539 Fall 2003Project Overview•Goal: For a given route, estimate the average and the standard deviation of drive time•Inputs: –Number of Stop Signs–Number of Stop Lights–Number of MilesData Collection•12 different routes (Input Vectors)•Each route driven 4 times to establish average and varianceModel Evaluation•Use 7 input vectors for training•Use 2 for comparison between models•Use final 3 for comparison with competition:–2 Experienced Drivers–YahooMaps!Neural Network Models Tried•Radial Basis Network (Type I and Type II)•Various Multi-Layer Perceptron Models•Single Linear Perceptron Model–Best Results.–Agreed with intuitive structure of the physical systemModel Details:•Average Drive Time: Straight forward linear activation function was used:tavg = x1*w1 + x2*w2 + x3*w3 + wb•Drive Time Variation: Based on statistics definition of total variation of inputs in series the squared standard deviation is used:ResultsPractical Insight•The weights express the time cost associated with each stoplight, stop sign, and mile.•For example: Each stop sign adds ~ 20 seconds to the overall drive time of a given routetavg = x1*w1 + x2*w2 + x3*w3 + wbPractical Insight?•For the case of standard deviation of drive time, some of the values were negative. Since its squared, this implies an imaginary standard deviation?Conclusions•The data set was large enough for the results to be pretty good for the average drive time.•The results for the variation were not too far off, but not great. •Recommendations: If further improvement of the model is desired...–More data points for each vector should be acquired to improve the accuracy of the experimentally measured standard deviation for each route–More training vectors should be
View Full Document