The proliferation of consumer global position service (GPS) devices in conjunction with an increasing number of municipalities employing traffic monitoring systems, has opened the door for advanced traffic control as well as personalized route planning. The detailed information about non-highway traffic is particularly useful in european cities where traffic is dominated by small, local roads.

In this research, we consider information about the length of traffic queues on highly busy intersections in Moscow, Russia. We will build a model for each congested area that combines,

• historical information, by learning a density for the queue-length given the time;
• current information, by learning a graphical model that relates the congestion between intersections.

Using the combined model, we will demonstrate potential applications such as finding the travel-time optimal route for a commuter, as well as dynamic control of traffic lights in a network.

In the past few years, the number of global position system (GPS) devices and dynamic mapping applications has skyrocketed. This includes standalone GPS devices as well as smart-phones and portable web-browsing devices. Such devices often give the current location, as well as current traffic information to the user. In some instances, these devices can be programmed to avoid traffic, by dynamically changing the route as traffic is encountered.

Traffic information is most often available for highway and main-flow roads, however municipalities have begun placing traffic monitoring systems at more intersections within cities, with the intended purpose of enforcing traffic laws, as well as acquiring data for traffic control and planning.

Combining the increasingly detailed traffic information with location-aware devices can lead to powerful tools for the commuter and city alike. For instance, as already mentioned, some GPS devices will dynamically adjust routes according to current local traffic conditions, however, these devices do not take into account the predicatable traffic behavior of the route at the outset of the journey. Information about current traffic in other areas of the city is also not accounted for, even though it may effect the commuter later in the trip.

In this project, we propose to model the traffic behavior of intersections for a portion of the road system in Moscow, Russia. From these models, we will demonstrate the optimal route for a commuter, given his current location and time, by exploiting the current traffic conditions in the entire section of the city in addition to historical information about the intersection.

Specifically, we will learn two models for each intersection,

• a density function for each intersection that models the traffic conditions for a given time/day;
• a graphical model that relates the traffic conditions of each intersection to each other for a given time.

By combining these models, we will be able to take advatage of both historical and current information for each intersection. We will use the data provided by Yandex.ru to learn the models. This data includes "queue-length" information for each intersection, the time and date, and geographical location.

We will verify the models using standard machine learning verification techniques. We have a particular advantage with regards to verification in that we can continue to collect data throughout the duration of the project.

Once the models are learned and verified, we will implement an algorithm to compute the best commuter route. Such an algorithm will take the users current location, destination location, current traffic conditions, date/time, and the learned models as parameters, to produce a route that minimizes the travel time.