01698034

Component

This paper examines the use of a committee of forecasting models in various transportation applications. The methods used in forming a committee of predictors are surveyed and discussed. As a result, a criteria list for choosing the final committee components is set. Four different approaches for combining the forecasts of individual models are tested: (1) Consensus forecasts (2) Bagging (3) Reinforcement Learning and (4) Bayesian Estimation. Each of the four techniques is tested on forecasting flight departures from O’Hare International Airport, and on the number of entries into Howard metro rail station in Chicago. The performance of the individual prediction models is compared to the performance of the committee. The performance across committees is also examined. The results confirm the improvement in accuracy when using a committee of forecasters. For both data sets, Reinforcement Learning (RL) and Consensus forecasts had similar performance although RL was slightly better, and both methods outperformed the other approaches. By analyzing the error distributions the authors were able to trace the change in the prediction uncertainty when moving from individual components to committees. Finally, the committees were able to limit the outlier frequency and generate predictions with a small error distribution standard deviation unlike the individual components.

This paper was sponsored by TRB committee ABJ70 Standing Committee on Artificial Intelligence and Advanced Computing Applications.

19-05306

English

0000-0002-8443-8928

7p

2019

Transportation Research Board 98th Annual Meeting

Digital/other

Figures; References

Accuracy; Arrivals and departures; Forecasting; Machine learning; Methodology; Rail transit stations; Travel demand

Chicago O'Hare International Airport; Howard Station

Chicago (Illinois)

Aviation; Planning and Forecasting; Public Transportation

Transportation Research Board Annual Meeting 2019 Paper #19-05306

TRIS, TRB, ATRI

Dec 7 2018 9:44AM