Organic Control of Traffic Lights
Organic Control of Traffic Lights Autonomic and Trusted Computing (ATC-08), Oslo, Norway June 24th, 2008
H. Prothmann*, F. Rochner**, S. Tomforde**, J. Branke*, C. Müller-Schloer**, H. Schmeck*
*
**
www.kit.edu
Outline 1. 2.
3.
4.
2
Motivation Motivation An organic architecture for traffic control Overview Important architectural aspects Experimental evaluation Test scenario Results Summary and outlook
Congestion in Los Angeles
m$ congestion cost
9325 1452 43
% of daily travel is congested
m liter increased fuel consumption
72
hours annual delay per traveler
[D. Schrank, T. Lomax: THE 2007 URBAN MOBILITY REPORT] 3
Organic traffic lights Better traffic light control can help to ease congestion problems! Why “organic” traffic lights? Adaptation: Adapt autonomously to the environment Long term changes Short term fluctuations, incidents Learning: Learn new control strategies, … Autonomy and self-optimisation Limit necessary manual intervention for setup and maintenance Decentralised collaboration among neighbouring traffic nodes
Organic traffic lights
4
Outline 1. 2.
3.
4.
5
Motivation An organic architecture for traffic control Overview Important architectural aspects Experimental evaluation Test scenario Results Summary and outlook
Layer 2
Observer
LCS Layer 1
detector data
Traffic Light Controller (TLC)
signal settings
System under Observation/Control
6
User EA
Layer 2
Simulator
Definition of system objectives Off-line parameter optimisation • Evolutionary Algorithm (EA) evolves TLC parameters • Simulation-based evaluation
Layer 1
objectives (LOS, …)
On-line parameter selection • Observer monitors traffic • Learning Classifier System (LCS) selects TLC parameters and learns rule quality
SuOC
Architecture
Control of traffic signals • Industry-standard TLC • Fixed-time • Traffic-responsive • Parameters determine performance
Layer 1: On-line parameter selection
Observer Turning A B veh/h 500 750
… …
A
Observer avg. delay = 24
B
Match set [MS] 7
Condition Action Pred. *420,530+ *700,800+*…+ : TLC05 26 *420,510+ *740,910+*…+ : TLC04 18 *100,250+ *350,430+*…+ : TLC07 23 *290,510+ *700,980+*…+ : TLC04 22 *420,530+ *700,800+*…+ : TLC05 26 *420,510+ *740,910+*…+ : TLC04 18 *290,510+ *700,980+*…+ : TLC04 22
Action set [AS]
Rule base
TLC04
Prediction update
*420,510+ *740,910+*…+ : TLC04 18 *290,510+ *700,980+*…+ : TLC04 22
Prediction array TLC04 TLC05 20 26
LCS
Layer 2: Off-line parameter optimisation B
[420,530] [700,800] : TLC05 26 [100,250] [350,430] : TLC07 24 <empty>
add new rule
Rule base
A
MS
Turning A B veh/h 100 1050
initialise
Rule creation in “classical” LCS Covering (Create matching rules randomly) Random crossover / mutation of existing rules Environment used for evaluation
Simulator
… in OTC EA optimises TLC parameters for observed situation Simulation-based evaluation
Population
EA Parents
create
Offspring
Layer 2 8
Undesired side-effect: Weak competition among rules
Condition predicate of turning B
Condition predicate of turning A
Off-line optimisation takes time, but LCS must react immediately If no matching rules exist Select “closest” rule, copy, widen condition Activate rule generation
Condition predicate of turning A
Layer 1: LCS modifications (1)
C1
C2 Condition predicate of turning B
9
Layer 1: LCS modifications (2)
Match set [MS]
[220,530+ *300,800+*…+ : TLC05 20 *120,510+ *600,990+*…+ : TLC04 18 *420,490+ *770,820+*…+ : TLC07 17 …
*220,530+ *300,800+*…+ : TLC05 20 *120,510+ *600,990+*…+ : TLC04 18 [420,500] [750,820+*…+ : TLC07 17
Action set [AS]
Rule base
*500, 750, …+
[420,500] [750,820+*…+ : TLC07 17
Prediction array TLC04 TLC05 TLC07 18 20 17
LCS
while (< d distinct actions in MS and rule with distinct action in proximity) add widened copy to MS if (widened copy in AS) add widened copy to rule base
10
Outline 1. 2.
3.
4.
11
Motivation An organic architecture for traffic control Overview Important architectural aspects Experimental evaluation Test scenario Results Summary and outlook
Test scenario Intersections located at Hamburg, Germany
K7
Traffic demands: From traffic census Reference: Fixed-time controller provided by traffic engineer Criterion: Avg. delay time
f d f tT
tT
12
t
t
t
T set of intersection’s turnings ft flow for turning t dt avg. delay for turning t
K3
Results (K7)
Improvement by organic approach 13
Day 1
Day 2
Day 3
10%
12%
12%
Results (K3)
Improvement by organic approach 14
Day 1
Day 2
Day 3
6%
8%
8%
Outline 1. 2.
3.
4.
15
Motivation An organic architecture for traffic control Overview Important architectural aspects Experimental evaluation Test scenario Results Summary and outlook
Summary and outlook Summary Traffic control is an interesting application for Organic Computing. The presented Organic Traffic Control application adapts autonomously to its environment, learns new control strategies when necessary, and thereby limits manual intervention and effort for setup and maintenance. Evaluation in a realistic scenario showed promising results.
Outlook Establish progressive signal systems by decentralised collaboration
16
H. Prothmann*, F. Rochner**, S. Tomforde**, J. Branke*, C. Müller-Schloer**, H. Schmeck*
*
**
www.kit.edu
Outline 1. 2.
3.
4.
2
Motivation Motivation An organic architecture for traffic control Overview Important architectural aspects Experimental evaluation Test scenario Results Summary and outlook
Congestion in Los Angeles
m$ congestion cost
9325 1452 43
% of daily travel is congested
m liter increased fuel consumption
72
hours annual delay per traveler
[D. Schrank, T. Lomax: THE 2007 URBAN MOBILITY REPORT] 3
Organic traffic lights Better traffic light control can help to ease congestion problems! Why “organic” traffic lights? Adaptation: Adapt autonomously to the environment Long term changes Short term fluctuations, incidents Learning: Learn new control strategies, … Autonomy and self-optimisation Limit necessary manual intervention for setup and maintenance Decentralised collaboration among neighbouring traffic nodes
Organic traffic lights
4
Outline 1. 2.
3.
4.
5
Motivation An organic architecture for traffic control Overview Important architectural aspects Experimental evaluation Test scenario Results Summary and outlook
Layer 2
Observer
LCS Layer 1
detector data
Traffic Light Controller (TLC)
signal settings
System under Observation/Control
6
User EA
Layer 2
Simulator
Definition of system objectives Off-line parameter optimisation • Evolutionary Algorithm (EA) evolves TLC parameters • Simulation-based evaluation
Layer 1
objectives (LOS, …)
On-line parameter selection • Observer monitors traffic • Learning Classifier System (LCS) selects TLC parameters and learns rule quality
SuOC
Architecture
Control of traffic signals • Industry-standard TLC • Fixed-time • Traffic-responsive • Parameters determine performance
Layer 1: On-line parameter selection
Observer Turning A B veh/h 500 750
… …
A
Observer avg. delay = 24
B
Match set [MS] 7
Condition Action Pred. *420,530+ *700,800+*…+ : TLC05 26 *420,510+ *740,910+*…+ : TLC04 18 *100,250+ *350,430+*…+ : TLC07 23 *290,510+ *700,980+*…+ : TLC04 22 *420,530+ *700,800+*…+ : TLC05 26 *420,510+ *740,910+*…+ : TLC04 18 *290,510+ *700,980+*…+ : TLC04 22
Action set [AS]
Rule base
TLC04
Prediction update
*420,510+ *740,910+*…+ : TLC04 18 *290,510+ *700,980+*…+ : TLC04 22
Prediction array TLC04 TLC05 20 26
LCS
Layer 2: Off-line parameter optimisation B
[420,530] [700,800] : TLC05 26 [100,250] [350,430] : TLC07 24 <empty>
add new rule
Rule base
A
MS
Turning A B veh/h 100 1050
initialise
Rule creation in “classical” LCS Covering (Create matching rules randomly) Random crossover / mutation of existing rules Environment used for evaluation
Simulator
… in OTC EA optimises TLC parameters for observed situation Simulation-based evaluation
Population
EA Parents
create
Offspring
Layer 2 8
Undesired side-effect: Weak competition among rules
Condition predicate of turning B
Condition predicate of turning A
Off-line optimisation takes time, but LCS must react immediately If no matching rules exist Select “closest” rule, copy, widen condition Activate rule generation
Condition predicate of turning A
Layer 1: LCS modifications (1)
C1
C2 Condition predicate of turning B
9
Layer 1: LCS modifications (2)
Match set [MS]
[220,530+ *300,800+*…+ : TLC05 20 *120,510+ *600,990+*…+ : TLC04 18 *420,490+ *770,820+*…+ : TLC07 17 …
*220,530+ *300,800+*…+ : TLC05 20 *120,510+ *600,990+*…+ : TLC04 18 [420,500] [750,820+*…+ : TLC07 17
Action set [AS]
Rule base
*500, 750, …+
[420,500] [750,820+*…+ : TLC07 17
Prediction array TLC04 TLC05 TLC07 18 20 17
LCS
while (< d distinct actions in MS and rule with distinct action in proximity) add widened copy to MS if (widened copy in AS) add widened copy to rule base
10
Outline 1. 2.
3.
4.
11
Motivation An organic architecture for traffic control Overview Important architectural aspects Experimental evaluation Test scenario Results Summary and outlook
Test scenario Intersections located at Hamburg, Germany
K7
Traffic demands: From traffic census Reference: Fixed-time controller provided by traffic engineer Criterion: Avg. delay time
f d f tT
tT
12
t
t
t
T set of intersection’s turnings ft flow for turning t dt avg. delay for turning t
K3
Results (K7)
Improvement by organic approach 13
Day 1
Day 2
Day 3
10%
12%
12%
Results (K3)
Improvement by organic approach 14
Day 1
Day 2
Day 3
6%
8%
8%
Outline 1. 2.
3.
4.
15
Motivation An organic architecture for traffic control Overview Important architectural aspects Experimental evaluation Test scenario Results Summary and outlook
Summary and outlook Summary Traffic control is an interesting application for Organic Computing. The presented Organic Traffic Control application adapts autonomously to its environment, learns new control strategies when necessary, and thereby limits manual intervention and effort for setup and maintenance. Evaluation in a realistic scenario showed promising results.
Outlook Establish progressive signal systems by decentralised collaboration
16
