1. Adaptive Traffic Control
Brian Dong & Graham Roff

2. Motivation Traffic Jams: gridlock at rush hour, urban sprawl
Current implementation: Fixed-cycle or even fixed-interval, environment insensitive design
Optimizing every intersection is not practical: requires collection of data at each intersection in the city to analyze traffic patterns at all hours of day
Would be nice if a traffic light controller could figure this out all on its own…

3. Design Challenges
Goal: Develop an adaptive algorithm and embedded traffic controller that can optimize traffic flow without any a priori information other than which lanes can be green simultaneously.
Algorithm complexity: how to weigh the various factors/parameters affecting the intersection lanes (previous wait times, traffic volume, current cars waiting)
Simulation: how to create a reasonably accurate representation of cars arriving at an intersection

4. System Design
Modular design: intersection module contains lane modules, each containing a hardware interface
Simulator communicates by triggering the hardware (car sensor) and responding to hardware state changes (light changed)
Intersection module controls changing state (lights) of intersection
Decisions based on state “metric delta”

5. System Diagram

6. Algorithm Design Decisions based on state “metric” for each lane
A given state causes a change of metric over time:
The metric delta is evaluated over every possible next state
State with best reduction in metric is chosen

7. Algorithm Challenges Time: how to determine t in the above equations
Lanes go green until idle or max timeout expires
Red light time for other lanes is minimum time light will be red for based on light concurrency
Simultaneous lanes: multiple lanes may be green at the same time, and light changes are asynchronous
Algorithm must handle switching the first idle lane vs. waiting for all green lanes to go idle before switching to new state

8. Simulator Design
Needs to represent the bursty nature of traffic: groups of cars arrive sequentially at an intersection, there are time delays between these groups/bursts.
Solution: Use two Gaussian random variables of different means and variances.
First variable determines the time delta between cars during a burst
Second variable determines the time delta between the bursts
Use a uniformly distributed random variable to determine when the burst occurs
Interface with TLC controller: Generate test vectors using Matlab and compile into SystemC. Simulator module actuates the TLC hardware sensors based on the delta-queue created in Matlab

9. Example and Demonstration

10. Results
Three different control algorithms were implemented for comparison
Fixed time
Fixed cycle
“Optimum”
Measured the mean and variance of the waiting time, and the percentage of stopped cars
A basic four-way intersection with eight lanes (four straight and four left) was used for analysis
Number of traffic scenarios examined, including:
All lanes equally busy
All straight lanes busy
North/south straight lanes busy

11. Results - 2
The results for the equally busy intersection are shown below:
Fixed Interval
Fixed Cycle
Optimum
Wait time mean
67.0
52.2
42.1
Wait time variance
42.7
36.2
35.2
Stopped cars
94%
92%
Fixed interval is worst performer, followed by fixed cycle
Optimum intersection performs the best
Relatively small difference between fixed cycle and optimum – however optimum requires no additional state change configuration
Plot of lane queue length versus time is also generated to examine algorithm behavior

12. Results - 3

13. Future Directions Basic principle is proven, further testing remains
Support for pedestrians, emergency vehicles, etc…is not included as these are not relevant to algorithm efficiency
Further improvement could be obtained by allowing intersection communication – beyond the scope of this project

14. Conclusions
Complex problem: Traffic light control is more complicated than one might expect
Many possible algorithms: Impossible to declare one algorithm “optimal” as there is no real definition of optimum
Our general “metric” based approach allows easy tweaking of details to optimize the system for a given algorithm
For a simple waiting time evaluation this algorithm performs the best of the three tested and is simplest to configure
Testing: Biggest concern in a production environment is testing – any solution needs to be tested across every possible scenario

15. Questions?