1 of 27

Machine Learning (ML) �What it can bring to travel modelling

Peter Vovsha, Principal Scientist, Mobility Simulation

peter.vovsha@bentley.com

1

2 of 27

Three Focused Aspects

ML as replacement for / enhancement of logit models

Logical and controlled model elasticity as current stumbling block

ML as behavior analysis tool

Uncovering non-linear combinations of behavioral variables & effects

ML ideas adapted for model calibration

Systematic approach to calibration of travel model system

2

3 of 27

ML as replacement for / enhancement of logit models

Logit

ML

3

4 of 27

Single choice example to compare different models

Auto ownership model:

0, 1, 2, 3+ cars

6 model applications:

4 ML + 2 MNL (highest probability chosen and microsimulation)

Model assessment criteria:

Predictive power (individual hit & aggregate shares)

Behavioral insights

Policy sensitivity

4

5 of 27

Single model performance / slightly in favor of ML

5

6 of 27

Sensitivity tests / ML pitfalls

Scenario 1 🡪 Better transit service: all IVTs halved
Expectation: car ownership would decrease

Scenario 2 🡪 Worse transit service: all IVTs doubled
Expectation: car ownership would increase

Transit travel time changed for the entire region

Desired elasticities guaranteed by the model structure

Only MNL passed these tests

6

7 of 27

Practical pros and cons of logit vs. ML for single model

Logit models

Machine Learning

Transparency

Controllable elasticity

Adjustment of constants

What the data says

Individual hit

Only applicable to a single model w/specific targets

Hybridization is attractive direction:

Mixed MNL-NN – the best of both worlds?

More details:

Vyas, G., P. Vovsha (2020) Assessment of Machine Learning methods versus logit models for travel modelling in practice.

Presented at the 99th TRB Annual Meeting, Washington, D.C.

7

8 of 27

ML as auxiliary analysis tool

8

9 of 27

Important technical differences

No clear feedback from model estimation to specification
Strong assumptions made a priori as a price for controlled elasticities
Data is used to fine-tune parameters

Logit models separate model specification and estimation

Model specification and training are closely intertwined
Data is used to shape the model
Less a priori assumptions and hence less controlled elasticities

ML methods

9

10 of 27

Examples of typical non-linear effects in travel behaviour

Identification of thresholds is essential ��ML was extremely helpful!

More details:

Wang, S., G. Vyas, P. Vovsha (2018) Interlinkage between Trip Chaining and Mode Choice.

Presented at the 97th Annual Meeting of the Transportation Research Board.

10

11 of 27

ML ideas adapted for model system calibration

Travel model

Activity frequency form HTS

Big data: trip departure profiles

Big data: OD trip distribution

Traffic counts

Transit ridership

11

12 of 27

Known limitations of conventional model estimation

Sub-models are estimated one at a time based on the “perfect” observed prior choices

However, they are applied conditional upon each other

LOS i.e., accessibilities are assumed fixed and external

However, in model application LOS is equilibrated

Aggregate data such as traffic or transit counts or “big data” cannot be utilized

However, matching these types of data is important in practice

Automated calibration resolves all 3 issues!

12

13 of 27

ML power to infer relationships across the entire model system

Special methods for adjustment of parameters

Backward propagation

Association between parameters and targets

Transportation model system as a whole

Neural network

Gradient method for adjustment of parameters

Backward propagation

Transportation models one at time

Special methods for adjustment of parameters

Targets corresponding to parameters directly

13

14 of 27

AGENT as the platform:

Assemble virtually any travel demand model structure from 4-step to ABM
Transparent access to a flexible UI for travel demand modelling
Maintain or version different model structures in parallel
Upgrade and advance models over time with new features
Automated calibration of entire model system

14

15 of 27

Application Examples

	1. MAG (Phoenix, AZ) Weekend ABM	2. Lima, OH ABM
Calibration targets	Weekend activity rates from HTS Big data O-D tables (AirSage data) Traffic counts by time periods	Big data O-D tables (StreetLight data) Traffic counts by time periods
Initial model state	Default values of coefficients from the weekday model	Previously estimated & calibrated model using conventional methods

Key calibration logic:

Data

Targets

Sub-models

Coefficients

15

16 of 27

MAG: calibration instrumentation (1/4)

Work location

Auto ownership

Tour frequency

Tour destination

Tour mode choice

Stop location

Trip mode choice

Weekend traffic counts

Tour frequency constants

Tour TOD

Trip departure

Weekend AirSage time-of-day distribution

Weekend AirSage OD

Stop frequency

Weekend activity frequency (from

HTS)

Stop frequency constants

16

17 of 27

MAG: calibration instrumentation (2/4)

Work location

Auto ownership

Tour frequency

Tour destination

Tour mode choice

Stop location

Trip mode choice

Weekend traffic counts

Tour TOD

Trip departure

Time period constants by tour purpose

Time period constants by trip purpose

Weekend AirSage time-of-day distribution

Weekend AirSage OD

Stop frequency

Weekend activity frequency (from literature)

17

18 of 27

MAG: calibration instrumentation (3/4)

Work location

Auto ownership

Tour frequency

Tour destination

Tour mode choice

Stop location

Trip mode choice

Auto ownership constants

Tour frequency constants

Dispersion coefficient
Intra-zonal preference
Attraction rates

Dispersion coefficient
Intra-zonal preference
Attraction rates

Tour TOD

Trip departure

Stop frequency

Stop frequency constants

Weekend traffic counts

Weekend AirSage time-of-day distribution

Weekend AirSage OD

Weekend activity frequency (from literature)

18

19 of 27

MAG: calibration instrumentation (4/4)

Work location

Auto ownership

Tour frequency

Tour destination

Tour mode choice

Stop location

Trip mode choice

Auto ownership constants

Tour frequency constants

Dispersion coefficient
Intra-zonal preference
Attraction rates

Dispersion coefficient
Intra-zonal preference
Attraction rates

Auto mode constants

Auto mode constants

Tour TOD

Trip departure

Time period constants

Time period constants

Stop frequency

Stop frequency constants

Weekend traffic counts

Weekend AirSage time-of-day distribution

Weekend AirSage OD

Weekend activity frequency (from literature)

19

20 of 27

MAG Weekend ABM: Matching link traffic counts

20

21 of 27

Lima ABM: Validation to big data

R²: 0.57, %RMSE: 259%

R²: 0.92, %RMSE: 115%

21

22 of 27

Validation to traffic counts: % RMSE reduction/increase

Importance of Big Data	Importance of traffic counts
	0.0		0.5		1.0
	MD	PM	MD	PM	MD	PM
0.0	-	-			-16%	-5%
0.5					-18%	-1%
1.0	-21%	15%	-19%	-1%	-22%	-3%

Importance of Big Data	Importance of traffic counts
Importance of Big Data	0.0	0.5	1.0
0.0	-		10%
0.5			-50%
1.0	-56%	-59%	-53%

Validation to big O-D data: % RMSE reduction/increase

How big data and traffic counts work together (Lima)

Big data and traffic counts mostly help each other and combined calibration is beneficial

22

23 of 27

Automated calibration summary:

Improve model calibration and validation results
Minimize risks and costs associated with costly trial-and-error approaches to calibration
Keep travel demand models up-to-date across mobility changes
Incorporate all mobility data sources including big data
AGENT works seamlessly with EMME and CUBE

AGENT

HTS

OD Data

Traffic Counts

Fare Card Data

Ridership Data

More details:

Vyas, G., P. Vovsha (2023) ML in applied travel modelling. Presented at the Innovations in Transportation Applications Planning Conference, Indianapolis, IN

Hasnat, M., G. Vyas, P. Vovsha (2023) Disaggregate estimation vs. Aggregate Calibration. Presented at the ETC conference, Milan.

23

24 of 27

Conclusions on Machine Learning

Fascinating spectrum of new directions for improvement of travel models

Replacement of logit models by ML has challenges for forecasting out-of-band scenarios

Uncovering non-linear effects is a low-hanging fruit with general purpose tools

Systematic model system calibration proved extremely beneficial, available to practitioners now

AGENT is now the recommended demand modelling system for all EMME and CUBE users

24

25 of 27

Relevant publications for more details

Vyas, G., P. Vovsha (2023) ML in applied travel modelling. Presented at the Innovations in Transportation Applications Planning Conference, Indianapolis, IN
Hasnat, M., G. Vyas, P. Vovsha (2023) Disaggregate estimation vs. Aggregate Calibration. Presented at the ETC conference, Milan.
Vyas, G., P. Vovsha (2020) Assessment of Machine Learning methods versus logit models for travel modelling in practice. Presented at the 99th TRB Annual Meeting, Washington, D.C.
Wang, S., G. Vyas, P. Vovsha (2018) Interlinkage between Trip Chaining and Mode Choice. Presented at the 97th Annual Meeting of the Transportation Research Board.

25

26 of 27

Thank you!

Peter Vovsha, Principal Scientist, Mobility Simulation

peter.vovsha@bentley.com

26

27 of 27

Our Colors

Bentley Black

RGB 26, 26, 26

WEB 1A1A1A

Bentley Green

RGB 68, 214, 44

WEB 44D62C

Pantone 802 C

Bentley Grey

CMYK 0, 0, 0, 30

RGB 188, 190, 192

WEB bcbec0

Bentley Grey

WEB d1d3d4

Bentley Grey

WEB e6e7e8

Bentley Med Blue

CMYK 81, 38, 0, 0

RGB 20, 134, 200

WEB 1485c7

Bentley Orange

CMYK 0, 70, 97, 0

RGB 243, 112, 36

WEB f37024

Bentley White

RGB 255, 255, 255

WEB FFFFFF

CMYK | RGB | WEB | PMS

The Bentley color palette has been carefully crafted and has evolved to convey the ideas �of a stable, professional, technical, and �modern software company. The use of the colors represents the overall brand and maintains a sense of consistency throughout Bentley's marketing.

27