Inrix Overview

1
The Science of Prediction
How Next Generation Traffic Services Will Impact
Business Dr. Oliver Downs, Chief Scientist
Location Intelligence ConferenceApril 4, 2006
2
Traffic Congestion Impacts Business
ProductivityFor Urban Areas Traffic Increases
Represent Major Business Hurdle

• During the last decade in the nations 68 largest
  urban areas
• Time wasted due to traffic rose from 1.9 to 4.5
  billion hours
• Heavily congested roads increased from 14 to 36
• Congested roads increased from 21 to 28
• Uncongested roads decreased from 65 to 36

• Congestion has real costs and affects business
  productivity
• Hard costs
• Lost labor time from employees spending hours in
  traffic
• Extra time for pick-up and delivery/reduced
  production time
• Extra vehicles to meet just-in-time demands of
  customers and scheduling problems caused by
  longer delivery times
• Soft costs
• Business credibility

3
The Evolution
Of Traffic Information

• Traffic Information is evolving to greater levels
  of customer value

• Prediction Relevance
• System-wide analysis of traffic
• Useful for planning, re-routing, analysis and
  decisions

• Flow Data
• Real-time speed information
• Useful for digital consumption

4
What
We Do
Aggregate traffic-related content from public and
private sources
Enhance real-time data using proprietary error
detection and correction. Utilize sophisticated
Bayesian modeling for prediction.
Distribute to customers via XML services
5
Traffic Service
Real-Time Flow Prediction

• Features
• Estimates of traffic flow patterns every 15
  minutes for up to one year in the future
• Amount of time expected for congestion to start
  or clear at Inrix Traffic Segments
• Drive time predictions for key routes in all
  metropolitan markets
• Proprietary error detection and correction of
  individual real-time traffic sensors
  significantly increases the quality of the flow
  reporting

Inrix indicates the most likely duration of a
current jammed segment of the highway, making it
easy to calculate route times
Inrix provides the time until the adjacent road
segment will most likely become jammed.
6
Bayesian Reasoning
And Prediction

• The Bayesian approach to predictive modeling
  allows us to learn the parameters ? of an
  appropriate probabilistic generative model from a
  set of data,

To do so, we define variables in the model that
correspond to the unknown parameters ?, assign
priors to these variables based on our background
knowledge e and use Bayes' rule to update our
beliefs about these parameters given observed
data
Bayes rule
We can then average over the learned
distributions of ? to make predictions
7
Prediction and Forecasting
Leveraging Bayesian Networks

• A Bayesian network is a graphical model that can
  be used to uncover causal relationships between
  a large number of variables
• Inrix utilizes sophisticated Bayesian modeling
  incorporating metadata attributes that we
  would expect to influence the observed sensor
  data both now and some time in the future

• Inrix Metadata
• Current/ Historical Traffic
• Time of Day
• Day of Week

• Weather
• Sporting Events
• Season
• School Schedules

• Holiday Status
• Incident/ Construction Reports
• Express Lane Direction

8
Prediction and Forecasting
Rules Underlying the Causal Relationships

• The set of conditionally independent causal
  relationships combine to describe a series of
  rules that are probabilistically predictive of a
  given outcome of the target variables, a decision
  tree

• Each end leaf of the tree, describes a particular
  rule that a current traffic situation may follow,
  and a probability distribution over the predicted
  outcome
• Color 109 False and Color 118 not False and
  Black Start 24 'None' and Pct b15 75 not
  '50-75' and Color 132 False and Black Start 144
  not 'None' and Color 119 False and Pct b15
  132 '30-50

9
Bayesian vs. Marginal Models
Why Only Bayesian Models Provide Accurate
Forecasts

• Bayesian modeling incorporating contextual
  metadata provides significantly higher
  quality/accuracy than simple marginal models
  based upon historical traffic

Rolling Stones Concert, 7.30pm at Key Area in
Seattle 10/30/2005. Effect of Bayesian
incorporation of metadata into forecast 48 hours
ahead, for 7pm 10/30/2005.
10
Example Customer
Use Cases

• Providing intra day and predicted travel times
  and travel speeds for key routes
• Delivering time dependent, traffic influenced
  routing and dynamic rerouting with turn-by-turn
  navigation
• Generating 2D and 3D traffic speed maps showing
  real-time, predicted forecasted traffic hot
  spots

• Illustrating expected times for congestion to
  clear
• Providing 1-day, 5-day or 10-day metropolitan
  traffic forecasts
• Analyzing traffic congestion bottlenecks for
  transportation and site planning
• Highlighting real-time traffic alerts on a map
• Developing highly personalized traffic reports
  and SMS alerts

11
(No Transcript)