4.1. Traffic Assignment Procedure

Along with a network data model, traffic assignment is the most technically challenging portion to develop in a modeling platform, especially if you want it to be FAST. In AequilibraE, we aim to make it as fast as possible, without making it overly complex to use, develop and maintain (we know how subjective complex is).

Note

AequilibraE has had efficient multi-threaded All-or-Nothing (AoN) assignment for a while, but since the Method-of-Successive-Averages, Frank-Wolfe, Conjugate-Frank-Wolfe and Biconjugate-Frank-Wolfe are new in the software, it should take some time for these implementations to reach full maturity.

4.1.1. Performing traffic assignment

For a comprehensive use case for the traffic assignment module, please see the Traffic assignment section of the use cases page.

4.1.1.1. Traffic Assignment Class

Traffic assignment is organized within a object new to version 0.6.1 that includes a small list of member variables which should be populated by the user, providing a complete specification of the assignment procedure:

  • classes: List of objects Traffic class , each of which are a completely specified traffic class

  • vdf: The Volume delay function (VDF) to be used

  • vdf_parameters: The parameters to be used in the volume delay function, other than volume, capacity and free flow time

  • time_field: The field of the graph that corresponds to free-flow travel time. The procedure will collect this information from the graph associated with the first traffic class provided, but will check if all graphs have the same information on free-flow travel time

  • capacity_field: The field of the graph that corresponds to link capacity. The procedure will collect this information from the graph associated with the first traffic class provided, but will check if all graphs have the same information on free-flow travel time

  • algorithm: The assignment algorithm to be used. e.g. “all-or-nothing” or “bfw”

Assignment parameters such as maximum number of iterations and target relative gap come from the global software parameters, that can be set using the Parameters module .

There are also some strict technical requirements for formulating the multi-class equilibrium assignment as a contrained convex optimization problem, as we have implemented it. These requirements are loosely listed in Technical requirements .

If you want to see the assignment log on your terminal during the assignment, please look in the Logging section of the use cases.

To begin building the assignment it is easy:

from aequilibrae.paths import TrafficAssignment

assig = TrafficAssignment()

4.1.1.1.1. Volume Delay Function

For now, the only available VDF function in AequilibraE is the BPR, but more functions will be added as needed/requested/possible.

\(CongestedTime_{i} = FreeFlowTime_{i} * (1 + \alpha * (\frac{Volume_{i}}{Capacity_{i}})^\beta)\)

Setting the volume delay function is one of the first things you should do after instantiating an assignment problem in AequilibraE, and it is as simple as:

assig.set_vdf('BPR')

The implementation of the VDF functions in AequilibraE is written in Cython and fully multi-threaded, and therefore descent methods that may evaluate such function multiple times per iteration should not become unecessarily slow, especially in modern multi-core systems.

4.1.1.2. Traffic class

The Traffic class object holds all the information pertaining to a specific traffic class to be assigned. There are three pieces of information that are required in the composition of this class:

  • graph - It is the Graph object corresponding to that particular traffic class/ mode

  • matrix - It is the AequilibraE matrix with the demand for that traffic class, but which can have an arbitrary number of user-classes, setup as different layers of the matrix object (see the Setting multiple user classes before assignment

  • pce - The passenger-car equivalent is the standard way of modelling multi-class traffic assignment equilibrium in a consistent manner (see [4] for the technical detail), and it is set to 1 by default. If the pce for a certain class should be different than one, one can make a quick method call.

Example:

tc = TrafficClass(graph_car, matrix_car)

tc2 = TrafficClass(graph_truck, matrix_truck)
tc2.set_pce(1.9)

To add traffic classes to the assignment instance it is just a matter of making a method call:

assig.set_classes([tc, tc2])

4.1.1.3. setting VDF Parameters

Parameters for VDF functions can be passed as a fixed value to use for all links, or as graph fields. As it is the case for the travel time and capacity fields, VDF parameters need to be consistent across all graphs.

Because AequilibraE supports different parameters for each link, its implementation is the most general possible while still preserving the desired properties for multi-class assignment, but the user needs to provide individual values for each link OR a single value for the entire network.

Setting the VDF parameters should be done AFTER setting the VDF function of choice and adding traffic classes to the assignment, or it will fail.

To choose a field that exists in the graph, we just pass the parameters as follows:

assig.set_vdf_parameters({"alpha": "alphas", "beta": "betas"})

To pass global values, it is simply a matter of doing the following:

assig.set_vdf_parameters({"alpha": 0.15, "beta": 4})

4.1.1.4. Setting final parameters

There are still three parameters missing for the assignment.

  • Capacity field

  • Travel time field

  • Equilibrium algorithm to use

assig.set_capacity_field("capacity")
assig.set_time_field("free_flow_time")
assig.set_algorithm(algorithm)

Finally, one can execute assignment:

assig.execute()

Convergence criteria is discussed below.

4.1.2. Multi-class Equilibrium assignment

By introducing equilibrium assignment [1] with as many algorithms as we have, it makes sense to also introduce multi-class assignment, adding to the pre-existing capability of assigning multiple user-classes at once. However, multi-class equilibrium assignments have strict technical requirements and different equilibrium algorithms have slightly different resource requirements.

Note

Our implementations of the conjudate and Biconjugate-Frank-Wolfe methods should be inherently proportional [6], but we have not yet carried the appropriate testing that would be required for an empirical proof

4.1.2.1. Cost function

It is currently not possible to use custom cost functions for assignment, and the only cost function available to be minimized is simply travel time.

4.1.2.2. Technical requirements

This documentation is not intended to discuss in detail the mathematical requirements of multi-class traffic assignment, which can be found discussed in detail on Zill et all.

A few requirements, however, need to be made clear.

  • All traffic classes shall have identical free-flow travel times throughout the network

  • Each class shall have an unique Passenger Car Equivalency (PCE) factor

  • Volume delay functions shall be monotonically increasing. Well behaved functions are always something we are after

For the conjugate and Biconjugate Frank-Wolfe algorithms it is also necessary that the VDFs are differentiable.

4.1.2.3. Convergence criteria

Convergence in AequilibraE is measured solely in terms of relative gap, which is a somewhat old recommendation [5], but it is still the most used measure in practice, and is detailed below.

\(RelGap = \frac{\sum_{a}V_{a}^{*}*C_{a} - \sum_{a}V_{a}^{AoN}*C_{a}}{\sum_{a}V_{a}^{*}*C_{a}}\)

The algorithm’s two stop criteria currently used are the maximum number of iterations and the target Relative Gap, as specified above. These two parameters are collected directly from the Parameter File, described in detail in the Assignment section.

In order to override the parameter file values, one can set the assignment object member variables directly before execution.

assig.max_iter = 250
assig.rgap_target = 0.0001

4.1.2.4. Algorithms available

All algorithms have been implemented as a single software class, as the differences between them are simply the step direction and step size after each iteration of all-or-nothing assignment, as shown in the table below

Algorithm

Step direction

Step Size

Method of Successive Averages

All-or-Nothing assignment (AoN)

function of the iteration number

Frank-Wolfe

All-or-Nothing assignment

Optimal value derived from Wardrop’s principle

Conjugate Frank-Wolfe

Conjugate direction (Current and previous AoN)

Optimal value derived from Wardrop’s principle

Biconjugate Frank-Wolfe

Biconjugate direction (Current and two previous AoN)

Optimal value derived from Wardrop’s principle

4.1.2.4.1. Method of Successive Averages

This algorithm has been included largely for hystorical reasons, and we see very little reason to use it. Yet, it has been implemented with the appropriate computation of relative gap computation and supports all the analysis features available.

4.1.2.4.2. Frank-Wolfe (FW)

The implementation of Frank-Wolfe in AequilibraE is extremely simple from an implementation point of view, as we use a generic optimizer from SciPy as an engine for the line search, and it is a standard implementation of the algorithm introduced by LeBlanc in 1975 [2].

4.1.2.4.3. Conjugate Frank-Wolfe

The conjugate direction algorithm was introduced in 2013 [3], which is quite recent if you consider that the Frank-Wolfe algorithm was first applied in the early 1970’s, and it was introduced at the same as its Biconjugate evolution, so it was born outdated.

4.1.2.4.4. Biconjugate Frank-Wolfe

The Biconjugate Frank-Wolfe algorithm is currently the fastest converging link- based traffic assignment algorithm used in practice, and it is the recommended algorithm for AequilibraE users. Due to its need for previous iteration data, it requires more memory during runtime, but very large networks should still fit nicely in systems with 16Gb of RAM.

4.1.2.5. Implementation details & tricks

A few implementation details and tricks are worth mentioning not because it is needed to use the software, but because they were things we grappled with during implementation, and it would be a shame not register it for those looking to implement their own variations of this algorithm or to slight change it for their own purposes.

  • The relative gap is computed with the cost used to compute the All-or-Nothing portion of the iteration, and although the literature on this is obvious, we took some time to realize that we should re-compute the travel costs only AFTER checking for convergence.

  • In some instances, Frank-Wolfe is extremely unstable during the first iterations on assignment, resulting on numerical errors on our line search. We found that setting the step size to the corresponding MSA value (1/ current iteration) resulted in the problem quickly becoming stable and moving towards a state where the line search started working properly. This technique was generalized to the conjugate and biconjugate Frank-Wolfe algorithms.

4.1.2.5.1. Opportunities for multi-threading

Most multi-threading opportunities have already been taken advantage of during the implementation of the All-or-Nothing portion of the assignment. However, the optimization engine using for line search, as well as a few functions from NumPy could still be paralellized for maximum performance on system with high number of cores, such as the latest Threadripper CPUs. These numpy functions are the following:

  • np.sum

  • np.power

  • np.fill

A few NumPy operations have already been parallelized, and can be seen on a file called parallel_numpy.pyx if you are curious to look at.

Most of the gains of going back to Cython to paralelize these functions came from making in-place computation using previously existing arrays, as the instantiation of large NumPy arrays can be computationally expensive.

4.1.2.5.2. References

[1] Wardrop J. G. (1952) “Some theoretical aspects of road traffic research.” Proc. Inst. Civil Eng. 1 Part II, pp.325-378.

[2] LeBlanc L. J., Morlok E. K. and Pierskalla W. P. (1975) “An efficient approach to solving the road network equilibrium traffic assignment problem” Transpn Res. 9, 309-318.

[3] Maria Mitradjieva and Per Olov Lindberg (2013)”The Stiff Is Moving—Conjugate Direction Frank-Wolfe Methods with Applications to Traffic Assignment”, Mitradjieva and Lindberg

[4] Zill, J., Camargo, P., Veitch, T., Daisy,N. (2019) “Toll Choice and Stochastic User Equilibrium: Ticking All the Boxes”, Transportation Research Record, Vol 2673, Issue 4 Zill et. all

[5] Rose, G., Daskin, M., Koppelman, F. (1988) “An examination of convergence error in equilibrium traffic assignment models”, Transportation Res. B, Vol 22 Issue 4, PP 261-274 Rose, Daskin and Koppelman

[6] Florian, M., Morosan, C.D. (2014) “On uniqueness and proportionality in multi-class equilibrium assignment”, Transportation Research Part B, Volume 70, pg 261-274 Florian and Morosan

4.1.3. Handling the network

The other important topic when dealing with multi-class assignment is to have a single consistent handling of networks, as in the end there is only physical network being handled, regardless of access differences to each mode (e.g. truck lanes, High-Occupancy Lanes, etc.). This handling is often done with something called a super-network.

4.1.3.1. Super-network

We deal with a super-network by having all classes with the same links in their sub-graphs, but assigning b_node identical to a_node for all links whenever a link is not available for a certain user class. It is slightly less efficient when we are computing shortest paths, but a LOT more efficient when we are aggregating flows.

The use of the AequilibraE project and its built-in methods to build graphs ensure that all graphs will be built in a consistent manner and multi-class assignment is possible.

4.1.4. Numerical Study

Similar to other complex algorthms that handle a large amount of data through complex computations, traffic assignment procedures can always be subject to at least one very reasonable question: Are the results right?

For this reason, we have used all equilibrium traffic assignment algorithms available in AequilibraE to solve standard instances used in academia for comparing algorithm results, some of which have are available with highly converged solutions (~1e-14): https://github.com/bstabler/TransportationNetworks/

4.1.4.1. Sioux Falls

Network has:

  • Links: 76

  • Nodes: 24

  • Zones: 24

Sioux Falls MSA 500 iterations Sioux Falls Frank-Wolfe 500 iterations Sioux Falls Conjugate Frank-Wolfe 500 iterations Sioux Falls Biconjugate Frank-Wolfe 500 iterations

4.1.4.2. Anaheim

Network has:

  • Links: 914

  • Nodes: 416

  • Zones: 38

Anaheim MSA 500 iterations Anaheim Frank-Wolfe 500 iterations Anaheim Conjugate Frank-Wolfe 500 iterations Anaheim Biconjugate Frank-Wolfe 500 iterations

4.1.4.3. Winnipeg

Network has:

  • Links: 914

  • Nodes: 416

  • Zones: 38

Winnipeg MSA 500 iterations Winnipeg Frank-Wolfe 500 iterations Winnipeg Conjugate Frank-Wolfe 500 iterations Winnipeg Biconjugate Frank-Wolfe 500 iterations

The results for Winnipeg do not seem extremely good when compared to a highly, but we believe posting its results would suggest deeper investigation by one of our users :-),

4.1.4.4. Barcelona

Network has:

  • Links: 2,522

  • Nodes: 1,020

  • Zones: 110

Barcelona MSA 500 iterations Barcelona Frank-Wolfe 500 iterations Barcelona Conjugate Frank-Wolfe 500 iterations Barcelona Biconjugate Frank-Wolfe 500 iterations

4.1.4.5. Chicago Regional

Network has:

  • Links: 39,018

  • Nodes: 12,982

  • Zones: 1,790

Chicago MSA 500 iterations Chicago Frank-Wolfe 500 iterations Chicago Conjugate Frank-Wolfe 500 iterations Chicago Biconjugate Frank-Wolfe 500 iterations

4.1.5. Convergence Study

Besides validating the final results from the algorithms, we have also compared how well they converge for the largest instance we have tested (Chicago Regional), as that instance has a comparable size to real-world models.

Algorithm convergence comparison

Not surprinsingly, one can see that Frank-Wolfe far outperforms the Method of Successive Averages for a number of iterations larger than 25, and is capable of reaching 1.0e-04 just after 800 iterations, while MSA is still at 3.5e-4 even after 1,000 iterations.

The actual show, however, is left for the Biconjugate Frank-Wolfe implementation, which delivers a relative gap of under 1.0e-04 in under 200 iterations, and a relative gap of under 1.0e-05 in just over 700 iterations.

This convergence capability, allied to its computational performance described below suggest that AequilibraE is ready to be used in large real-world applications.

4.1.6. Computational performance

Running on a Thinkpad X1 extreme equipped with a 6 cores 8750H CPU and 32Gb of 2667Hz RAM, AequilibraE performed 1,000 iterations of Frank-Wolfe assignment on the Chicago Network in just under 46 minutes, while Biconjugate Frank Wolfe takes just under 47 minutes.

During this process, the sustained CPU clock fluctuated between 3.05 and 3.2GHz due to the laptop’s thermal constraints, suggesting that performance in modern desktops would be better

4.1.7. Noteworthy items

Note

The biggest opportunity for performance in AequilibraE right now it to apply network contraction hierarchies to the building of the graph, but that is still a long-term goal