Feedback on Feedback: CAMPO’s Findings from Testing Various Feedback Approaches
description
Transcript of Feedback on Feedback: CAMPO’s Findings from Testing Various Feedback Approaches
Feedback on Feedback: CAMPO’s Findings from Testing Various Feedback ApproachesTRB Applications ConferenceMay 11, 2011Session 18B
Feedback on Feedback: CAMPO’s Findings from Testing Various Feedback Approaches
Kevin LancasterCapital Area Metropolitan Planning Organization
Jonathan AvnerWilbur Smith Associates
Karen LorenziniTexas Transportation Institute
Feedback on Feedback: CAMPO’s Findings from Testing Various Feedback Approaches• Why Feedback?• What Did We Test?• What Did We Find?• Where To Next?
The CAMPO Model• Capital Area Metropolitan Planning Organization• Five Counties Encompassing the Austin – Round
Rock, Texas Metropolitan Area• Auto, Truck, Fixed Route and Bus Transit,
Bicycles, and Pedestrians• Generalized Cost Assignment Including Tolled
Facilities• 1413 Internal, 49 External Traffic Analysis Zones• 16628 (2035), 14480 (2005) Links• 11575 (2035), 10443 (2005) Nodes
Why Feedback?
• Recommended by previous peer reviews
• Intuitively justified because inputs into earlier steps of the model could be inconsistent with the model outputs at later stages
Original CAMPO ProcessTraditional Four-Step Sequential Process
Trip Generation (TRIPCAL5)
Trip Distribution (ATOM2)
Highway Assignment
Assigned Volumes/Congested Speeds
Speeds are not an input to Trip Generation
For Trip Distribution, speeds representative of 24-hour free-flow are used as input.
Speeds representative of 24-hour free-flow are used as input for the first Traffic Assignment iteration.
Mode Choice
Transit Assignment
Ridership/Boardings
How Did We Approach Feedback?• We Need to Decide:
What gets fed back? What convergence criteria to use?
• How We Decided: Research literature Research State of Practice
(TMIP and other Texas MPOs)
Various Common ApproachesDifferent Possible
ApproachesOptions for What
Gets Fed BackTypical Convergence
Measures• Naïve (Direct)• Fictive Costs• Methods of
Successive Averages (MSA)
• Constant Weight Methods
•Link Time•Link Volumes
(converted to time)
•Trips•Skims
•Absolute or Percentage Differences
– Typically system-wide measures
•Total Misplaced Flows– Typically trip matrices or
link volumes•Root Mean Square Error
(RMSE)–Typically skims or trip
tables•GEH Statistic– Empirical formula typically
applied to link volumes
What CAMPO Tested – Feedback Approaches
Approach What Gets Fed Back
• Method of Successive Averages (Caliper’s MSA Implementation)
•Link Volumes Processed into Time Values
• Constant Weight Method - 50 – 50 - 70 – 30 - 80 – 20
•Trip Tables Processed Prior to Assignment
What CAMPOTested
1ST MODEL RUN "LOOP 1"
TRIP GENERATIONTRIP DISTRIBUTION
MODE CHOICEASSIGNMENT (24-Hr and AM Pk Per Highway)
if MSA Method:Apply MSA formula to link volumes, derive
"MSA-times" to feedback
Convergence or Maximum Loops?
No
Yes Finalize Model Reports
ASSIGNMENT (24-Hr and AM Pk Per Highway)
SUBSEQUENT LOOPS
TRIP DISTRIBUTION (all steps)MODE CHOICE (all steps except Mkt Seg)
TRIP TABLES (24-Hr and AM Pk Per Highway)
Update Time Fields for SkimmingFeedback
if Constant Weights Method:For each of the 24-hr and AM Pk TRIP TABLES,
create a new trip table based on the Constant Weight factor: CW * (current loop trips) + (1-CW) * (previous
loop trips)
Feedback
if Constant Weights Method:
Feedback times resulting from assignment
24-Hour / Non-Work Trip
Purposesif MSA Method:
Apply MSA formula to link volumes, derive
"MSA-times" to feedback
if Constant Weights Method:
Feedback times resulting from assignment
2-Hour Peak / Work Trip Purposes
MSA Method Formula
What CAMPO Tested – Convergence Criteria• Aggregate
Total number of trips• Matrix Level
Trip and skim table changes• Link Level
GEH statistic Maximum link flow change
Feedback Report
Measures for Convergence Criteria• Total Number of Trips
Absolute value, percent change• Trip and Skim Table Changes
Percent RMSE, Percent Total Misplaced Flow
• Link Level Total link flow change,
maximum link flow change,GEH statistic
GEH Statistic
• What is it? Empirically-based, not true statistic test
Typically applied to link volumes
Invented in the 1970s
What Did We Find?
• For All Approaches, the Measures of Convergence We Tested Tended toward Stability
• Some Converged Fasterthan Others
Daily / 24-Hour MetricsPercent Change
Total Trips
Trip Table Change - % RMSE
0
50
100
150
200
250
1 2 3 4 5 6 7 8 9 10
Perc
ent R
MSE
Iteration
MSACW 50%CW 75%CW 80%
Skim Table Change - % RMSE
0
5
10
15
20
25
1 2 3 4 5 6 7 8 9 10
Perc
ent R
MSE
Iteration
MSACW 50%CW 75%CW 80%
Maximum Link Flow Difference
-0.014
-0.012
-0.01
-0.008
-0.006
-0.004
-0.002
0
0.002
0.004
1 2 3 4 5 6 7 8 9 10
Perc
ent
Iteration
MSACW 50%CW 75%CW 80%
0
5000
10000
15000
20000
25000
1 2 3 4 5 6 7 8 9 10
Max
Flo
w D
iffer
ence
Iteration
MSACW50CW75CW80
Daily / 24-Hour Metrics - GEH
0
5
10
15
20
25
30
35
40
45
50
1 2 3 4 5 6 7 8 9 10
% o
f Lin
ks G
EH >
5
Iteration
MSACW50CW75CW80
2-Hour / Peak Period MetricsPercent Change
Total Trips
Trip Table Change - % RMSE
Skim Table Change - % RMSE
Maximum Link Flow Difference
-0.014
-0.012
-0.01
-0.008
-0.006
-0.004
-0.002
0
0.002
0.004
0.006
1 2 3 4 5 6 7 8 9 10
Perc
ent
Iteration
MSACW 50%CW 75%CW 80%
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10
Perc
ent R
MSE
Iteration
MSACW 50%CW 75%CW 80%
0
50
100
150
200
250
300
350
1 2 3 4 5 6 7 8 9 10
Perc
ent R
MSE
Iteration
MSACW 50%CW 75%CW 80% Not
evaluated for peak period
Skim Change – % RMSE24-Hour Versus Peak Period
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10
Perc
ent R
MSE
Iteration
MSA 24-Hour
MSA Pk-Hour
Decision MatrixConsideration MSA Constant Weights
Performance • No significant differenceMathematical Rationale
• Mathematically proven to converge
•Empirically-demonstrated performance
Implementation and Maintenance
• Supported in TransCAD GISDK
• Coded using GISDK, not explicitly supported
State of Practice
• Seems that MSA might have a slight edge in the modeling community discussions
CAMPO’sChosen FeedbackMethod
1ST MODEL RUN "LOOP 1"
TRIP GENERATION
TRIP DISTRIBUTION
MODE CHOICE
ASSIGNMENT (24-Hr and AM Pk Hr Highway)
Feedback MSA-derived
24-Hour Timesfor
Non-Work Trip Purposes
Convergence or Maximum Loops?
No
Yes Finalize Model ReportsFeedback
ASSIGNMENT (24-Hr and AM Pk Hr Highway)
SUBSEQUENT LOOPS
TRIP DISTRIBUTION (all steps)
MODE CHOICE (all steps except Mkt Seg)
TRIP TABLES (24-Hr and AM Pk Hr Highway)
Feed
back
Update Time Fields for Skimming
Feedback MSA-derived
2-Hour Peak Timesfor
Work Trip Purposes
Feedback
Feed
back
ConvergenceCriteria:% RMSE ofSkim< .1
Lessons Learned
• Opportunity to address other inconsistencies
• For testing, run many, many iterations• Be cognizant of assignment convergence
issues that affect feedback• Running mode choice for each iteration
was appropriate (and defensible)• Run time was a factor in our
decisions
Where To Next?
• For the 2005 Model, CAMPO Continues to Investigate Project- and Link-Level Implications of Modeling with Feedback
• CAMPO is Working Toward a Time Period Modeling Approach for its 2010 Model
• Long-term, Investigating Incorporating Accessibility into Trip Generation, andLooping Feedback to Trip Generation
Feedback on Feedback: CAMPO’s Findings from Testing Various Feedback Approaches
For further information, please contact:Kevin Lancaster, Capital Area Metropolitan Planning Organization512/[email protected]
Jonathan Avner, Wilbur Smith Associates512/[email protected]
Karen Lorenzini, Texas Transportation Institute512/[email protected]