Predicting the effectiveness of commute reduction plans using neural networks

Abstract

Commute trip reduction plans are being implemented at an increasing number of worksites. In order to be able to structure the most effective plan for a specific worksite, it is necessary to understand the factors that determine commuter response to company incentives to change commuting habits. This study compares the predictive ability of neural networks compared to linear regression models in calculating the change in vehicle trip rate (VTR) at a given worksite over a year long travel plan period. Using a Los Angeles area dataset (n = 3439), linear regression and neural network models were constructed and optimized using input variables including worksite incentives, business type and number of years of incentives at the worksite. Significant differences(p=0.006, 0.007) in program effectiveness were discovered between the results of local authority worksites ([delta]VTR = -1.495) and both businesses ([delta]VTR = -0.986) and hospitals ([delta]VTR = -0.728). It was determined that the neural network (R² = 0.229) performed better than the linear regression models (R² = 0.038) when evaluated by R², representing an improvement of 0.014 on previous models.

(cont.) However, when measuring the ability of the models to predict within a certain interval around the output the linear regression models outperformed the neural network model by a factor of 35 percentage points. The lack of strong linear correlations between the inputs and the outputs of these models suggests that the most significant factors in creating successful transportation demand management programs are not currently being tracked. Given the statistically significant superior performance at local authority worksites it is suggested that more worksite demographics are tracked.