In this project, we propose a variational inference algorithm that tracks and predicts real-time traffic dynamics in a transportation network from an agent-based transportation model and multiple streaming data sources. To demonstrate the value of combining simulation modeling and big data in delivering travel information to drivers and promoting efficient driving, we will aggregate the noisy sensor network data from personal mobile phones on UB’s North Campus into transportation informatics such as available parking spaces in parking lots, and time and fuel consumption to find a parking space, suggest optimum trips to faculty, staff, and students through a real-time and interactive driving planner and promote fuel-efficient driving at UB’s North Campus. We will validate the variational inference algorithm and the driving planner with both synthesized and real data.
Many existing algorithms to track and predict real-time
traffic dynamics — vector ARIMA, state space,
neural network, and Bayesian network models — have
difficulties in coping with noisy and missing data,
making predictions in non-recurrent scenarios, and
explaining predictions in terms of agent trips. Our approach
to simulating agent trips in conjunction with traffic
dynamics captured by sensor networks overcomes
these difficulties by turning the problem of prediction into
that of searching compatible agent behaviors in a
probability space defined by the agent-based model. This
variational method of minimizing Bethe free energy is a
recent development in the field of machine learning.
There are two challenges in this project. One is that
mathematically, how can we effectively search in
the probability space for compatible agent behaviors
when enumerating the combinatorial state space is
computationally intractable? The other is that
practically, how can we combine an agent-based model and noisy
sensor network data when both model building and data
mining are time consuming?
This project integrates machine learning, big data, sensor networks, and agent-based transportation modeling to prototype an algorithm that combines the power of a model-driven approach with the power of big data, and promotes responsible driving by showing how different agent trips are associated with different travel time and fuel consumption.