USF Libraries
USF Digital Collections

Neural network application for predicting the impact of trip reduction strategies

MISSING IMAGE

Material Information

Title:
Neural network application for predicting the impact of trip reduction strategies
Physical Description:
Book
Language:
English
Creator:
University of South Florida. Center for Urban Transportation Research
Publisher:
Center for Urban Transportation Research (CUTR)
Place of Publication:
Tampa, Fla
Publication Date:

Subjects

Subjects / Keywords:
Ridesharing--United States   ( lcsh )
Neural networks (Computer science)   ( lcsh )
Commuting--United States   ( lcsh )
Traffic restraint (TRIS)   ( trt )
Travel behavior (TRIS)   ( trt )
Genre:
letter   ( marcgt )

Record Information

Source Institution:
University of South Florida Library
Holding Location:
University of South Florida
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
oclc - 39505884
usfldc doi - C01-00062
usfldc handle - c1.62
System ID:
SFS0032184:00001


This item is only available as the following downloads:


Full Text
xml version 1.0 encoding UTF-8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd
leader ntm 22 Ka 4500
controlfield tag 008 s flunnn| ||||ineng
datafield ind1 8 ind2 024
subfield code a C01-00062
035
(OCoLC) 39505884
040
FHM
049
FHmm
0 245
Neural network application for predicting the impact of trip reduction strategies
260
Tampa, Fla
b Center for Urban Transportation Research (CUTR)
c 1998 February
650
Ridesharing--United States
Neural networks (Computer science)
Commuting--United States
7
Traffic restraint (TRIS)
2 TRT
Travel behavior (TRIS)
TRT
710
University of South Florida. Center for Urban Transportation Research
720
Winters, Philip L.
1 773
t Center for Urban Transportation Research Publications [USF].
4 856
u http://digital.lib.usf.edu/?c1.62



PAGE 1

Neural Network Application for Predicting The Impact of Trip Reduction Strategies Philip L. Winters Francis Cleland Mark Burris Dr. Rafael Perez Michael Pietrzyk Center for Urban T ransportation Research University of South Florida 4202 E. Fowler Avenue, CUT 100 Tampa Florida 33620-5375 (813) 974-3120 February 1998 The opinions findings, and conclusions expressed in this publication are lhooe of the authol's and not necessarily those of the Rorida Department of Transpottalion or the U .S. Department of Transportation. This report has been prepared in cooperation with the State of Florida Department of Transportation and the U.S. Department of Transpot1ation, in partial fulflhment of HPR Study No. 0763, WP/ No. 0510763, State Job No. 99700-3337-119, Conlracl No. EH896, CUTR Account No. 21-17-189-LO, ent#Jed 'Neural Networl< Technology for the Evaluation of Trip Reduction Programs (Philip L Winters, CUTR; Dr. Rafael Perez, USF's Department of Computer Science and Engineering; and Michael Pietrzyk, CUTR. are the Cr:>-Principal ln'!Stigators). Project Manager is Ike Ubaka. FOOT Public Transit Office.

PAGE 2

TABLE OF CONTENTS I NEURAL NE1WORK APPLICATION FOR PREDICT ING THE IMPACT OF TRIP PREDUCTION STRATEGIES ... .. .... .. . ..... ................... . .. .... . .... .. .. ....... ......... ... ........ .... ... ...... 1 Background .. .... .. .... ...... ........... .. ............... .... .. .... ....... .... ....................................... ... ......... ......... 1 Project Objective ... ....... .... .. ... . .............. ... .... .... .... . .. . ........... .... .... ......... ... ...... ... ...... .......... . .... . 1 Project Overview ... ........... ........... ....... . .............. . . ....... ..... ......... . . . ...... ...... ...... ...... . . .... ....... ...... 2 What are Neura l Networks? .......... .... .................... ........... ....... .............. .... .. .............. .. ................... 2 Comparison of Neural Networks to Othe r Mode li ng Techniques .......... .. .. .. .. .. ...... .. .... .. .... .............. 3 II AlTERNATIVE MODELING PROCEDURES .... ....... .... .... ........ ..... .... ............ ............ ............ .... .... 4 RegreS$iOn Analysis ....................... .. ...... .. .................. .. .. .. ............................ .. ........ ... .. ................... 5 D iscrim i nant Ana ly s i s .... .................. .. .... .......... .. .... .. ...... .................................................. .... .. ...... .. ... 5 F HWA TOM Modei ....... ... ......... .... .... .... ....... ... .......................... ................................................ .... ... S Ill. MODE L -BUILD ING ACTIVI TIE S .... .............. ...... .... .............................. ....... ...... .......... .. ......... ... .... 6 Overview of Neura l Network Model Build i ng .... .... ...... ................ ............ .... .. .. .... ............ .... ........... 6 Data U se d for Mode l Building ....... ...... ............ .... ....... ........ .................................... ...... ................... 6 Criteria for Evaluating Model Perfonnanoe ............. ......... .. .................................. .... ....................... 7 Results of Alternative Modeling Procedures .. .......................... ..... . ..... ......... ...... ......... .... ........... 8 Final Model Building Results Us i ng On ly the SCAQMD Data .................. ...................... .... ........... 10 IV. FIELD TESTING ANN MODEL ..... .... .. ........................ ............... .... .... .. ........................................ 12 Approach to Field Testi ng the Mode l ................... ...... .................. ........ ........ ......... ........... ............... 12 Mode l lnoenli ves .. ... ................. .......... ............. .... . .. ................................ .. .... ..... ..... ... .......... ........... 12 V TECHNO LOGY TRANSFER ....................... .... ... ... .... ..... ........................ ...... ........ ........ ...... .... ..... 1 4 Software ... ............. ......... .................................................. .... .... ..... ......... .... ......... ........ ..... 14 Sample Tri p Reduction Plans ............... ......... ..... ..................... ............ .......................... ................ 18 VI. ADDITIONAL RESEARCH .............. .............. .......... .. .... .. .... .... ........... .... .. .... ....... ........ ........... ... 18 VII. CONC L USIONS ........ .... ....... .. ................ .. .................. .... ...... .................... .................... .. .......... ..... 21

PAGE 3

LIST OF TABLES AND FIGURES Table 1: AVR Change Range categories for Model Evaluation ... .. ....................... .. .. .. ........... ........ .. .... 9 Table 2: Acceptable Range Classification by Model for TOM Validation Data Set (N=432) ... . .... . . . . .... . 9 Table 3: linear Correlation of prediction and actual outpu1... .. ..... .. ..... ............ ... ...... .... .......... .... .. ........... ... 9 Table 4: Acceptable Classification by Models Using Full Data Set... ......................................................... 1 o Table 5: Acceptable Classification by Models Using Uncorrelated Data Set ............ .... ........... .. ....... ........ 11 Table 6: Acceptable Range Correct Classifocation by Final Models for TOM Model Validation Data Set (n=432) ............... ...................................................................................... 11 Table 7: Linear Correlation of Prediction and Actual Output TOM Model Validation Data Set (N=432) ........ ........ .... .. .............. ....................... ......... ............. .............. ... ...... .. ...... ... 11 Table 8: F inal Model Performance vs. Validation Data .... .......................................................................... 13 Table 9: Frequency of Incentives ............................................................ .... ..... ........... ..... ...... .................... 15 Table 10: Common Data Elements ... .. ................ ..... ...................... ........ .... ..... ... ....... .......... ....... ...... .... ... 16 Table 11: Employee Commute Information .... ................ ............................ .... ...... .... ................. ... .. ........... 19 Table 12: AVR Calculations ... ..................................... ... ...... .... ......... ................................. .... ............... .... 20 Table 13: Changes in AVR vs. Vehicles Per Employee Ratio .................. ......... .............. ...... .......... ...... .. .. 20 Figure 1: Typical Artificial Neuron ......................... ........... ................ .... ......... .... ................................ ........... 3

PAGE 4

Acknowledgements The project team would like t o extend thanks to the following individua l s who hel ped us with this project. Catherine Wasikowski and Dr. Waldo Lopez-Aqueres w ith th e South Coast Air Qua li ty Management District who agreed to share the data to allow us to deve lop the n eural net model. They a lso provided va luab le input and feedback du ring key points of the project Hong Kim and J ack Tsao of th e Southern Californ i a Assoc i ation of Governments who were helpfu l in obtaining the trave l i mpedance data for use in t h e model. Rita Hildebra nd a n d Nina Corson with the Pima Association of Governments, B i ll Kicksey a nd Phil Cummings with Maricopa County Tr i p Reduction Program, and Ra ndi A l cott with Valley Met r o i n Phoenix for allowing us to acquire the data from their respective trip reduction programs for testing purposes an d demonstrate the model software. Eric Schreffler, ESTC, for his review and comments on the development of other models in use and our approach. Finally, a special thanks to Ike Ubaka as project manager w ith Florida Departme nt of T ransportation for hi s patience and feedback.

PAGE 5

NEURAL NEI HORKAPPLICAT10N FOR PREDICTING THE IMPACT OF TRIP REDUCTION STRATEGIES Background Rising traffic congestion and air quality problems contribuled to federal state, and regional effor1s to reduce vehicle emissJons by requi ring latge employers to develop programs to reduce vehicle trips In areas with the worst air pollution, the program's goal. was to reduce driving-and pollution-by increasing the a\lelage numbe< o f employees in vehicles commuting to work (that is. average vehicle ridership or AVR). E m ployers were targeted by these regulations as employer policies such as work location, work schedule and parking policies strongly influence transportation mode choice decisions made by employees. In several of the major urban areas of the country (such as Los Angeles, Phoenix, SeatUe}, large employers with 100 or more employees were required by federal state or local regulation to detailed plans for influencing employee travel behavior in order to reduce air pollution and/or traffic congestion. Over the years, these metropolitan areas collected a large amount of data from these companies. lnfonnation was obtained that described different company site characteristics and the attemative modes of transportation available to the employees. The data also Included infonnation on th e types of financial and non-financial i n centives employers offered to employees. Employers provided infonnation on woll< schedules and alternative work arrangements such as telecommuting and compressed work weeks. They also collected information from employees on th e different modes of transportation selected by the employees and estimated the AVR Though areas such as Los Angeles had thousands of employer plans submitted under these regulations the regulators have had success i n developing models to predict changes i n AVR. P art of the reason for this rests with the complexity of the data. The Los Angeles area database, for example includes 62 different incentives that employers can select to increase A VR i n their work Some incentives are offered by relatively few employers Even when condensing the incentives into 28 calegories, the plans represented about 1 500 different combinations of incentives. At the same time, the cunent models (such as the FHWA TOM Model) are based on disaggregate data collected through relatively small samples of employers but augmented by employee surveys. Specifically, model predictions were not compared with actual resutts for any data that had not been used i n the model building process. Project Objective U nder this Florida Department ol Transportation (FOOT) Research Idea project, the project team of the Center for Urban Transportation Research (CUTR) and the Department of Computer Science and Engineering at the U n iversity of South Florida applied neural network technology to predict the impacts o f various trip reduction strategies on changes i n commute behavior. I n the eSfiy 1990s, COMSIS, a transportation consulting firm, was hired by the South Coast Air Quality Management District (SCAQMD) in the Los Angeles area to develop a tinear model to predict AVR They attempted to use the several thousand employer trip reduction plans to build the model. However the model did not perform to the satisfaction ol COMSIS. SCAQMD agreed to build a model using a data set developed by ARB for the California Air Resouroes Board (CARB) by COMSIS. The CARB model is a based model that used the resutts of surveys from only 45 employers. However, also i nduded data from 2,500 employees. Disaggregate employee data was not part of the AQMD data structure. Neural networks were selected because they can uncover the hidden relationships in the data from employers and the resulting change i n average vehicle ridership (AVR). The performance and selection o f the best model were based on compartng neural network output to actual AVR observations The neural network training (or learning) process allows the neural netwol1< model to predict the correct response to combinations of Input data values not previously seen by the network. The benefits of developing such a mode l are to streaml ine I

PAGE 6

development of trip reduction plans for employers, increase effectiveness of those plans. and provide a basis for consistent review by the regulating agencies. It should also improve efficiency by reducing regulatory slaff time in the review of employerfdeveloper trip reduction plans. Project Overview This project executive summary describeS neural network models, highl ights the efforts to build a model to predict changes in A VR, summarizes the development of the application, compares the neural network model performance with other analytical approaches, and summarizes the results of the field test The reader should review the four technical memoranda prepared as part of this project for more i nfonnation. Technical memorandum #1, "Regional Trip Reduction Databases." reports on the present state of trip reduction data management and analysis. Model inputs and outputs are Identified by reviewing several trip reduction ordinances. The technical memorandum also reviews previous attempts to develop a model including the TOM Model developed for FHWA, the California Air Resources Board TOM Model and the TOM Cost-Effectiveness Model developed for Pleasanton, California. Technical memorandum #2, "A Primer on Neural Networks in Transportation: Conoepts and Applications," discusses neural nelwoct Amendments of 1990 required large employers in these areas to submrt trip reduction plans on an annual basis. However, ECO was made voluntary in late 1995. These plans would have been the source of data that would have allowed many large urban areas to develop their own model or calibrate a national model. What are Neural Networtcs? Artificial neural nelwo
PAGE 7

To develop or "train' the model, the data set is usually divided into two groups-one group for training the network and another group for testing how well the networ1< has learned. A third independent data set is often reserved for validation Each train ing set of data is presented to the network. the output of the network differs from the com!CI output. the weights of individual networ1< nodes are changed. Training a neural network requires many cycles until the cumulative errors of all training sets are below an acceptable level as pre-defined by the neural networ1< builder. The lower this numbe< the better the networ1< is able to duplicate the associations between inputs and outputs in the training data. It is expected that once the network is able to dupficate the associations between inputs and outputs In the training data, will be able to produce correct outputs for input data not specifically included previously as part of the training data The training set of data uses an independent test data set against which to test predictions on a regular basis. Training is halted when the test perfolmance begins to degrade. Otherwise, the model may overfit the data. OVerfilling the training data occurs when the neural networ1< produces a nonlinear model that fits the training data pelfectly but fits the test data very poorly. inputs connection weights ---._;::--...... artificial neuron Figure 1 -Typical Artificial Neuron output Training a network using back propagation (the method used in this project) consists of finding the correct number of computational i n the network with the correct numerical values of the weights that connect these units so that the associations between input and output in an existing data set can be duplicated by the network Since each neuron implemen ts a mapping between its i nputs and output neural networks are capable of learning non-finear relationships that may exist in the data. This makes neural networks adaptable and especially useful in environments where the relationships between i nputs and outputs change over time. Comparison of Neural Networks to Other Modeling Techniques Neural networks deal with a broad range of problems. Artificial neural networks are known to be good at cl assification evaluation, optimization, decision-making, pattern recognition, behavior trend prediction, image analysis, filtering, and modeling control systems. There are some significant differences between expert systems and neural networks. Expert systems require that the relationships between the input data and the conclusions to be derived from that data be established before the expert system is buill The neural network needs the data from which can uncover the relationships, while the expert system needs the expert who has already learned those relationships. Anothel important difference can be found in the encoding of the data. Expert systelOS encode their knowledge in terms 3

PAGE 8

of rules, object and procedures. After training, neural nelwoll
PAGE 9

Regression Analysis An independent model was first crealed by means of factor analysis and s tepwise regression to provide a baseline of comparison of the ability of the neural network model to p r edict changes in AVR correcUy. Initial regressions suffered from multicollinearity withi n the data Since many independent variables were inten::orrelated, a possibUity exists that the coefficients resulting from model runs would not fully reflect the effects of each of the independent variables. The i nitial approach to eWminating the effects of the multicolfinearity was to run a factor analysis. Generally, factor analysis is used as a data reduction technique. The analytical procedure involves creating uncorrelated (orthogonaQ combinations of the i nitial dependent variables In common practice, the purpose of the analysis i s t o reduce a mass of variables to a reasonable number of elements (for example, 1 0) that the analyst can understand and eleplain. The stepwise regression was set to accept variables that significantly improved the model at an 85 percent confidence level. When th e analysis had been completed, the factors were then reconverted i nto the original component Independent variables. The conversion was made by multiplying the coefficients assigned by the regression model to the factors by the mabix of the factor loa d i ngs of the origina l variables The resulting equations predicted the change i n AVR. Linear and factor analyses were buitt using Statistical Analysis S)lstem (SAS). An attemative approach to reducing multicollinearity is to examine interoorrelations between the variables and to elimi nate variables unti l no h i g hly interoooelated combinations remain. Therefore, a correlation mabix of the variables was prepared, and policy-oriented variables with correlations more than 0.20 were e liminated from further estimations. This process also combined incentives into i ncentive groups< as described eaMier. Other variables (such as site descriptors, percentages of employees using modes or i n various jobs, etc.) remained i n the model. The variable set was reduced to a total of n reasonably uncorrelated" variables from the original set through examination of the correlation matrix. These variables were then used to produce both new neural network models and revised regression and discriminant (see below ) models. Stepwise procedures were used to build both the regression and discriminant models, and the neural net variable selection procedures were used for creating the neural net input sel Discriminant Analysis Comparing the neural network model' s pelfonnance against a categorical prediction modeling procedure was logical because CUTR already determined that models would be evaluated based on their ability to classify observations into categories The usual choice in transpoMation demand problems is to conduct a analysis. However, discrim in ant procedures, whi le methodologically leSS rigorous, provide the same types of res u lts and are much simpler to develop. The approach to the discriminant analysis model-building was sim ilar to the approach to the building of the regression model and used the same version of SAS a statistica l software package. Typically, the evaluation of a discriminant model is done by determining the percentage of observations correcUy classified in an independent test data sel In practice, th e results from test data sets tend closely to mirror the results from the data sets used to bu ild the models. The size of the initia l test data set (432 observations) was such that evaluation of classification patterns for anything but the overall sample was impractical. Results were reported for both th e test data set and for the base (or training) data set These results were reported because ij is Important to know not only overal l how well the model classifies results, but also whether there were a ny patterns o f misclassification, FHWA TOM Model Tes ting the neural network model' s pelforrnance against an existing trip reduction analytical tool was a sensible next step. The FHWA TOM Medel was selected because ij was the most commonly used tool available The FHWA TOM Model uses a pivot point procedure to estimate how changes in travel time or cost would affect mode shares. This model handles strategies other than changes i n time or cost as a system of look-up factors. The effecti'Jeness o f employer-based strategies is function of the TOM strategies used and employer participation in canying out those strategies. s

PAGE 10

The FHWA TOM Model requires that data be entered into the model that define the starting conditions, Including employer/site data on trips and moda l from site surveys. The primary inputs are either trip tables from the regional model or the mode spld of an area or employer. The next step gives the user flexibility to relate special conditions that may not propet1y reflect the starting data inp uts M. this point the user specifies the TOM strategies to be applied. The Model allows testing of any individual strategy. or as many as the user desires i n combination. The FHWA TOM Model separates TOM strategies into two groups: Area-wide Strategies or Employer-Based Strategies. Area-wide strategies are incenti\les provided by the public sector (such as high occupan cy vehicle Employer-based strategies are TOM strategies funded and/or carried out by individual businesses (such as subsidies). The approach to evaluating the FHWA TOM Model was to use a sample randomly extracted from the SCAQMD data set to compare models. The SCAQMO data conesponding to the descriptions for each level had to be converted i nto a form acceptable for input i nto the TOM Model to compare the neural net model with the FHWA TOM Model. Many of the SCAQMO data fields could be easily converted into inputs for the FHWA TOM Model A notable exception was how much time spent on the trip reduction program by the employee transportation coordinator Generally, SCAQMO data had to be aggregated for inclusion into the FHWA TOM Model. For the Employer Support Programs input screen, data for input were extracted for the carpool program, including; matching, employer-based matching, preferential parlnnected, feed fo1W31'd type. Neu ral networ1< mode l builders have applied these types of networ1
PAGE 11

PREDICT uses 70 percent of the data as the training set and 30 percent as the test although the network biJijder can change these values to any other proportions. The test set is an extract of the training set used while building the model to prevent overfitting Ovemtting the training data can occur when the neural netwcxk produces a nonlinear model that fils the trai ning data perfectly. but fils the test data very poorly. The goal is to fit the training and test data with about the same overall error. Therefore. the test data set is used to analyze the model's ability to interpolate the train/test data during training. Training i s haHed when the test perfonnance starts to degrade. The validation set is independent of the train/test set and typifies the data that will be seen by the model i n the outside The neural network software does not use the validation set in building the model. The SCAQMD database includes 62 different incentives that employers can select to increase AVR in their work srtes. One neural network was buiH where an 62 incentives were grouped into one category. Subsequent networks were buitt using more limrted incentive groups. As mentioned 9,096 were used to build !he networks and 432 to validate the networks after they were built Initially, the network parameter settings were tested to find optimal configuration for network performance. Criteria fior Evaluating Model Perfiormance The SCAQMD data contained many observations (more than 500) where employers had either a very large inaease or very large decrease in A VR. Nevertheless, the vast majority (almost 90 percent) of the data falls near .{).10 to +().20 change i n AVR. Models buin on prediction error minimization criteria may force their predictions to the middle of the range (that is predict little or no change in AVR). This approach causes the models to have much more accuracy in the middle ranges of AVR change than with the outliers (that is, large changes i n AVR). Preferably, a model should interpolate well over the entire range of the input values. The neural netwcxk software manual contains an example of exactly this type of srtuation: "Is the linear regression l ine shown in Figure [2) a good solution to this problem? The answer depends on how the model is used. The objective of linear regression is to minimize the sum squared error of the difference between the estimated and actual outputs. If that is what is required by system objectives this model does that. However, ff the purpose of the model is to interpolate well over the entire range of the input space, this model fails." (Neuralware documentation 1995)'. To get a more comprehensive evaluation of the network's effectiveness, determined that an examination of the network's abilrty to correctly classr1y each prediction into a range (or a category) of AVR change would be conducted The ranges were developed by partitioning the data into equal sized groups based on the number of plans that fell within each range (thatis, the value of the dependent variable) (See Table 1.) In the evaluation centers on the model's at);lrty to predict whether a given combination of site characteristics and incentives will produce a large i ncrease i n A VR, a small increase, virtually no increase, a small decrease, or a large deaease in AVR. Models were evaluated both through comparison of R (linear correlation) values of predicted and actual change in A VR and by their abilrty to dassr1y an observation into the comect group or into an adjacent group. This was termed acceptable' (as opposed to "correct") classification. (See Tables 2 & 3.) 1 NeutatHate ctoc:umentatiOn "Neura/'Noltt.Js Predict MaN..aar Introduction 1-8, Bulk11ng a Neural Net Model. NeuraiNare. h::.. 1995 7

PAGE 12

I 1.2 1 0 8 > 0.6 o. 1 0: 1 --' ...... --
PAGE 13

Table 1 AVR Change Range Categories for Model Evaluation AVR change catego-0 .03, <0.06 Small increase 0 .03 to 0 .059 CL 5: >0.00, <0 1 2 Moderate increase 0.06to0 .119 CL 6 : Any change more lhan +0.03 Large increase 0 .12 or more Ct. 7: Any change more lhan +{). 06 Table2 Acceptable Range Classification by Model for TOM Validation Data Set (N=432) MODEL Percent Neural network 53.1 Discriminant 54.6 Regression 49.1 FHWA TOM Model 39. 6 Table 3 Linear correlation of prediction and actual output TOM Model Validation Set (N=432) MODEL R Neural networl< 0.441 Regression 0.541 FHWA TOM Model 0 .032 9

PAGE 14

This led to Phase IV, where decided to examine the impacts of changing parameters on the neural netwoll< software, hoping to improve network pelfonnance. The foul1h phaSe was an attempt to vary a range of neural netwo11< settings in attempting best to understand how the neural nelwoll
PAGE 15

Table 5 Acceptable Classification by Models Using Unconelated Data Set Model I nputs R Overan CL 1 CL2 CL3 Cl.4 CI.S Cl. 6 Cl. 7 aooeptable classification Neural 19 .33 53% 44% 31% 46% 78% 86% 60% 23% Regress. 33 .49 55% 47% 32% 43% 73% 86% 65% 33% Discrim. 2 1 na 56% 54% 46% 65% 74% 68% 46% 35% TableS Acceptable Range Conec:t Classification bY Final Models for TOM Model Validation Data Set (N=432) MODEL Inputs Percent Neural network 16 54.2 Discriminant 23 58.1 Regression 31 50.2 FHWA TOM Model NIA 39.6 Table7 Linear Conelatlon of Prediction and Actual Output TOM Model Validation Data Set (N=432) MODEL R Neural network 0.312 Regression 0.544 FHWA TDM Model 0.032 The neural nelwoll< was. therefore, deemed to be the superior model built, although admittedly somewhat less able to outpeffonn the l i near procedures than initially anticipated None of the models had a significant change in the number of variables they used to make their p redictions although the identity of those variables did change some from model to model. As a final step in model building, a neural network-based classification approach was tested. The best resutt obtained was with a network with only superficial data transformation and no hidden The superficial data transformation creates just one transform (for eJ
PAGE 16

FIELD TESTING ANN MODEL Approach to Field Testing the Model The approach to evaluating the transferabilily of the ANN model was to use the Los Angeles-based ANN model to predict change in AVR using data from another city (that is, Tucson and Phoenix). However the Los Angeles-based ANN model did not pertorm as well with the data from other cities as did with data used from Los Angeles as a validation set (that is, Los Angeles data that wasn't used to build the Los Angeles-based ANN model). (See technical memorandum #3 for a description of the model building process.) The project team hypothesized that other variables not included in the data set could explain the differences between the urban areas. For example, Los Angeles' population density is about 5 times higher than PhoeniX and Tucson. Higher densities can provide for setVice with lower headways thus offering me
PAGE 17

The idea of d i scarding potentially substantial number of variab l es is sometimes hard to accept However, there are plausible reasons for their exclusion by the al golithm. It might seem unrealistic that on l y five TOM i ncentives can impact employee clloice of h ow to com mute Where are the mal1
PAGE 18

categories used in Phoenix and Tucson). The last column i n the table indicates which fields were used to build the ANN model based on the availability of the data from each dty and the ability to combine data fields (for example, guaranleed ride home programs using taxis and guaranteed ride home programs using fleet vehicles). For a complete description of the data elemenls in the LA data set, please refer to technical memorandum #3. TECHNOLOGY TRANSFER Software Trip reduction software (CUTR AVR) was developed as the result of this project Using the software: Employers and developers can reduce the costs for developing and implementing plans to reduce vehicle trips by streamlining the plan development and review process. Public agencies could Improve efficiency by reducing staff time i n the review of employer/developer tri p reduction plans. Analysis i n public agencies can develop consistent interpretations of trip reduction plans. Thc>Ugh the model was never intended to be an integral part of the tlanspo
PAGE 19

T able9 Fooquencyof lncenUves No. of P lans with No. of Plans without F inal Model lnoentive Incentive Incentive Rideshar e match ing 3,644 3,336 Inc l uded Guaranleed ride home 3 ,4136 3.494 Inc luded Alternative mode s u bsidies 3 ,227 3 744 Included Compressed work week 1 ,769 5 2 1 1 Inc luded High park ing costs f o r SOV 76 6 ,904 Inc luded Marketin g 4 ,459 2 521 Excluded Preferential parki n g 2 721 4 ,259 Exc luded Other services 2 ,655 4 ,325 Exc luded B i ke racks and l ockers 2 ,620 4 ,360 Excluded Flexible work anangements 1,914 5.066 E xcl u ded Showers & c lothing lockers 1 ,554 5 .426 Exclud ed Telecommuting 1,058 5,922 Exclu ded cafeteria, ATMs post office etc 1,01 9 5,961 Excluded Other services 920 6 ,060 Excluded Free meals 771 6,209 Excluded Other compressed work week 675 6 ,305 Excluded Child care service 597 6 ,383 Excluded Walk to work subsidies 454 6 526 Excluded cata log points 354 6 ,626 Excluded Service (unspecified) 320 6 ,660 Excluded G i ft certificates 304 6 ,676 Exclu ded Auto services 221 6 759 Excluded Additional time off with pay 153 6 8 2 7 Exclu ded Other non financia l incentives 127 6 853 Exclu ded Other facility i mprovements 117 6,863 Exc luded Other parking strategies 116 6,864 Excluded Company vanpools 98 6 ,882 Exc luded Facility i mprovements 33 6,947 Excluded Prize drawings 0 6,980 Exc l u d ed IS

PAGE 20

Ceniufor Urban RrSJtafdJ Table10 Common Data Elements Los Data Used to DATA ELEMENT Phoenix Tucson Angeles Build ANN Mode l Plan Sequence Indicator na Excluded Exclu ded Drive alone peroentage Excluded Excluded Excluded Included Mo t orcycle percentage Excluded na Excluded Excluded 2-Person carpool pet Excluded Excluded Excluded Excluded 3-Person carpool pet Excluded Excluded Excluded Included 4-Person carpool pet Excluded Excluded Excluded Exc luded 5-Person carpool pet. Excluded Excluded Excluded Exc luded 6+ Person carpoo l pet na na Excluded Excluded Vanpool percenlage Excluded Exc luded Excluded Excluded Buspool percentage na na Excluded Exc luded Excluded Excluded I ncluded Wa l k Excluded Excluded Excluded Excluded Bicycle Exc luded Excluded Excluded Included T eleoommute na na Excluded CurrentAVR Excluded Excluded Excluded Included TargetAVR n a na Excluded Standard Industrial Classification (SIC) Excluded Excluded Exclud ed No. of employees on Excluded Excluded Excluded Included No. of employees arriv ing between 6 and 10 a.m. Excluded Excluded Excluded Excluded Percent of administrative employees na na Excluded Percent of professional emp loyees na na Excluded Percent of technica l employees Excluded Excluded Exclu ded Percent of clerical employees na na Excluded Percent of skilled wOII
PAGE 21

C'nJUfoc lkberlr Los Data Used to DATA ELEMENT Phoenix Tucson Angeles Bui ld ANN Model Pet of employees w/5 to 10 min oommute na na Excluded Excluded Pet of "mployees w/10 to15 min. oommute Excluded Excluded Excluded Exc luded Pet of emp lo yees w/15 to 20 min oommute n a n a Excluded Excluded Pet of employees w/20 to 30 m in oommute na na Excluded Exc luded Pet of employees w/30 to 40 min. oommute Excluded Excluded Excluded Excluded Pet of employees w/40 to 60 min. oommute na n a Excluded Included Pet of employees w/60 to 90 min commute na na Excluded Included Pet of emp loyees w/90 to 12 0 min. oommute na na Excluded Inc l uded Pet of employees w/120+ min commute n a na Excluded Inc l uded Facility Improvements (unknown) Excluded Excluded Excluded Excluded Other facility improvements n a na Excluded Excluded Preferential par1
PAGE 22

Los Data Used to DATA ELEMENT Pfloenix Tucson Angeles BuiJd ANN Model In-house or regional ridematching system Excluded Excluded Excluded Included Any type of Teleoornmuting Program Excluded Excluded Excluded Excluded Any type of Compressed Work Week program Excluded Excluded Excluded Included Any type of Financiallncentive/Oisincentive Excluded Excluded Excl uded Included Pencentage of palldng reserved for poo l s na na Excluded Included = included in the final model Excluded = excluded in the final model Impedance values are grouped as a single variable representing the percent of emp/oy&eS commuting over 40 minutes one-way to work One of the features contained in 1he software is the ability to evaluate the impact of multiple employers (currently up to 1 00) and combine the r esults of 2 or more employer profiles. This feature will hel p regional agencies such as a transportation management organization evaluate the i mpact of the program i n a particular area or multiple employer sites At the same time, the data l'ecessary to run the model in a suJ>.area mode is not available. As explained in an technical memorandum, Flonda employers are not required to trip reduction plans so the data on number of large employers with given strategies is currently unknown. However, as part of the mobi l ity management process regional commuter assistance programs could be requested to collecl the data on a la rger scale. Sample Trip Reduc:tion Plans In add i tion sample plans based on 1he model were develOPed to allow employers and o1hers to estimate changes In AVR based on different mixes of key variables (for e>eample, employees at site current mode split, elc. ) ParUally as a result of meeting with the Arizona trip reduction program staff CUTR focused efforts on designing 1he output for sample trip reduction plans as guidance documents for employers and developers with 1hese pre-selected attributes Though the model was developed to predict the absolute change i n AVR, the Arizona TOM staff recommended thal1he results also be presented in other fonnats. As the attached sheels show CUTR added 1he Vehicle Employee Ratio (VER) 1hat shows the number of vehicles per 100 employees Also a t the suggestion of the Arizona TRP staff, CUTR estimated the number of vehicles reduced for that employe r For el
PAGE 23

Table 11 E m ployee Commute lnfonnation Table 13 shows that the same change in AVR can measure different impacts on vehicle travel and par1
PAGE 24

Table 12 AVR Calculations T able 1 3 Cha n ges in AVR vs. Vehicles Per Employee Ratio CHANGE IN AVR AVR VER -0.08 -0.03 0 0.03 0.06 1 .10 91 98 93 91 88 86 1 .50 07 70 68 67 65 64 Future research projects could seek to adapt the ANN trip reduction model to transpol1ation planning process in a similar manner to the FHWA TOM Model. The FHWA TOM a pM>I point model, modifies trip tables based on assumption s of individual strategies including employer participation based on size of the employer and regula!OI'f e n vironme n t I n the short tenn the ANN model could use the output of the mode model to estimate current AVR. Assuming a mix of employer sizes and a proportional dislribution of the reduction among zones, the model can ca lculate the number of vehicle trips reduced at the zo n a l level. Additional research could be undeltlken to evaluate the impacts of these assumptions. Assessing other means 20

PAGE 25

of gathering data to take advantage of the model' s sensitivity to variables such as the current AVR. the share of employees wi1h long distance oommutes and employer size could include oombining commercial databases and geographic information systems. The ANN model does have limitations. One of the l imitations of the ANN model is the lack of information on impacts of small employer programs. The data used to develop the ANN model is l imited to large employment sites due to the regulatory requirements in Los Angeles, Phoenix, and Tucson only applying to large employers. Another limitation is the use of dummy variables rather than disaete values. For example. the impact of finandal inoentives was based on whether incentives were offere9. not the amount of the incentive due to inconsistent reporting of the incentive (amount, number of employees, etc ) In general, the federal taxoode effectivejy the tax-free amount of subsidies to $15 to $21 per month in the tate 1980's and early 1990s. In 1992, the tax code was changed to allow empklyers to provide up to 560 per month tax-free to employees for and vanpool subsidies. CONCLUSIONS Based on this project. the ANN model has proven to predict an acceptable range of changes i n AVR and has proven to be transferatlje to another soo. The final products (software and sample plans) should be applicable to Florida Furthermore, the ANN model outperformed other analysis tools and is easier to use as evaluated by TOM professionals. Flllally the model provides a basis for helping transportation planners assess the impacts of employerbased TOM strategies on vehicle bips. 21