fast-trips is a Dynamic Transit Assignment tool written in Python and supplemented by code in C++. For more information about this visit the following links:
- Project Website: http://fast-trips.mtc.ca.gov/
- Full Technical Documentation (API): http://metropolitantransportationcommission.github.io/fast-trips/
Follow the steps below to setup up fast-trips:
- We suggest that you work from a Python 2.7 virtual environment in order to make sure you don't interfere with other python installations you can do this using the base virtenv package, conda, or using the Anaconda Navigator GUI. Using conda:
conda create -n fasttrips python=2.7 anacondasource activate fasttrips. - Your VirtEnv should include numpy, pandas, and transitfeed for reading GTFS Many data analytics Python distributions like Anaconda bundle numpy and pandas, but they can also be installed using the command
pip install <packagename>within the virtual environment. As a last resort, Windows users can also find binary package installers here. - Install Git and clone the fast-trips repository (https://github.com/MetropolitanTransportationCommission/fast-trips.git) to a local directory:
<fast-trips-dir>. If the user plans on making changes to the code, it is recommended that the repository be forked before cloning. - Switch to the
developbranch of the repository. - To build, in the fast-trips directory
<fast-trips-dir>, run the following in a command prompt:python setup.py develop build_ext --inplace. If compiling on Windows, install Microsoft Visual C++ Compiler for Python 2.7. On Linux, install the python-dev package. On Mac, using standard xcode command line tools / g++ works fine. Using the develop command prompt makes sure that changes in the package are propogated to the shell without having to re-install the package.
The input to fast-trips consists of:
- A Transit Network directory, including schedules, access, egress and transfer information, specified by the GTFS-Plus Data Standards Repository
- A Transit Demand directory, including persons, households and trips, specified by the Demand Data Standards Repository
- fast-trips Configuration, specified below
Configuration is specified in the following files:
This is a required python file and may be included in both the Transit Supply and Transit Demand input directories. If the same options are specified in both, then the version specified in the Transit Demand input directory will be used. (Two versions may be specified because some configuration options are more relevant to demand and some are more relevant to network inputs.)
The configuration files are parsed by python's ConfigParser module and therefore adhere to that format, with two possible sections: fasttrips and pathfinding. (See Network Example ) (See Demand Example )
| Option Name | Type | Default | Description |
|---|---|---|---|
bump_buffer |
float | 5 | Not really used yet. |
bump_one_at_a_time |
bool | False | |
capacity_constraint |
bool | False | Hard capacity constraint. When True, fasttrips forces everyone off overcapacity vehicles and disallows them from finding a new path using an overcapacity vehicle. |
create_skims |
bool | False | Not implemented yet. |
debug_num_trips |
int | -1 | If positive, will truncate the trip list to this length. |
debug_trace_only |
bool | False | If True, will only find paths and simulate the person ids specified in trace_person_ids. |
iterations |
int | 1 | Number of pathfinding iterations to run. |
number_of_processes |
int | 0 | Number of processes to use for path finding. |
output_passenger_trajectories |
bool | True | Write chosen passenger paths? TODO: deprecate. Why would you ever not do this? |
output_pathset_per_sim_iter |
bool | False | Output pathsets for each simulation iteration? If false, just outputs once per path-finding iteration. |
prepend_route_id_to_trip_id |
bool | False | This is for readability in debugging; if True, then route ids will be prepended to trip ids. |
simulation |
bool | True | After path-finding, should we choose paths and assign passengers? (Why would you ever not do this?) |
skim_start_time |
string | 5:00 | Not implemented yet. |
skim_end_time |
string | 10:00 | Not implemented yet. |
skip_person_ids |
string | 'None' | A list of person IDs to skip. |
trace_person_ids |
string | 'None' | A list of person IDs for whom to output verbose trace information. |
| Option Name | Type | Default | Description |
|---|---|---|---|
max_num_paths |
int | -1 | If positive, drops paths after this IF probability is less than min_path_probability |
min_path_probability |
float | 0.005 | Paths with probability less than this get dropped IF max_num_paths specified AND hit. |
min_transfer_penalty |
float | 1 | Minimum transfer penalty. Safeguard against having no transfer penalty which can result in terrible paths with excessive transfers. |
overlap_scale_parameter |
float | 1 | Scale parameter for overlap path size variable. |
overlap_split_transit |
bool | False | For overlap calcs, split transit leg into component legs (A to E becauses A-B-C-D-E) |
overlap_variable |
string | 'count' | The variable upon which to base the overlap path size variable. Can be one of None, count, distance, time. |
pathfinding_type |
string | 'stochastic' | Pathfinding method. Can be stochastic, deterministic, or file. |
stochastic_dispersion |
float | 1.0 | Stochastic dispersion parameter. TODO: document this further. |
stochastic_max_stop_process_count |
int | -1 | In path-finding, how many times should we process a stop during labeling? Specify -1 for no max. |
stochastic_pathset_size |
int | 1000 | In path-finding, how many paths (not necessarily unique) determine a pathset? |
time_window |
float | 30 | In path-finding, the max time a passenger would wait at a stop. |
user_class_function |
string | 'generic_user_class' | A function to generate a user class string given a user record. |
The path size overlap penalty is formulated by Ramming and discussed in Hoogendoorn-Lanser et al. (see References ).
When the pathsize overlap is penalized (pathfinding overlap_variable is not None), then the following equation is used to calculate the path size overlap penalty:
Where
- i is the path alternative for individual n
- Γi is the set of legs of path alternative i
- la is the value of the
overlap_variablefor leg a. So it is either 1, the distance or the time of leg a depending of ifoverlap_scale_parameteriscount,distanceortime, respectively. - *Li is the total sum of the
overlap_variableover all legs la that make up path alternative i - *Cin is the choice set of path alternatives for individual n that overlap with alternative i
- γ is the
overlap_scale_parameter - δai = 1 and δaj = 0 ∀ j ≠ i
From Hoogendoor-Lanser et al.:
Consequently, if leg a for alternative i is unique, then [the denominator is equal to 1] and the path size contribution of leg a is equal to its proportional length la/Li. If leg la is also used by alternative j, then the contribution of leg la to path size PSi is smaller than la/Li. If γ = 0 or if routes i and j have equal length, then the contribution of leg a to PSi is equal to la/2Li. If γ > 0 and routes i and j differ in length, then the contribution of leg a to PSi depends on the ratio of Li to Lj. If route i is longer than route j and γ > 1, then the contribution of leg a to PSi is larger than la/2Li; otherwise, the contribution is smaller than la/2Li. If γ > 1 in the exponential path size formulation, then long routes are penalized in favor of short routes. The use of parameter γ is questionable if overlapping routes have more or less equal length and should therefore be set to 0. Overlap between those alternatives should not affect their choice probabilities differently. The degree to which long routes should be penalized might be determined by estimating γ. If γ is not estimated, then an educated guess with respect to γ should be made. To this end, differences in route length between alternatives in a choice set should be considered.
This is an optional python file in the Transit Demand input directory containing functions that are evaluated.
This could be used to programmatically define user classes based on person, household and/or trip attributes.
To use a function in this file, specify it in the pathfinding configuration as the user_class_function.
(See Example )
TBD
Sample input files have been provided in <fast-trips-dir>\Examples\test_network to test the setup and also assist with the creation of new fast-trips runs. The input files include network files created from a small hypothetical test network and also example transit demand data.
To quickly test the setup, run fast-trips on sample input using the following steps:
- Add
<fast-trips-dir>to thePYTHONPATHenvironment variable in Advanced system settings. - Run
\scripts\runAllTests.batfrom within<fast-trips-dir>in a command prompt. This will run several "preset" parameter combinations. The user can alternatively run each parameter combination individually using the commands listed in the batch file. Details about the test runs are provided in subsequent sections. Output files from running fast-trips with the sample input data provided can be found in theoutputdirectory.
A hypothetical 5-zone test network was developed to help code development. It has a total of three transit routes (one rail and two bus) with two or three stops each. There are also two park-and-ride (PnR) locations.
Transit vehicles commence at 3:00 PM and continue until 6:00 PM. There are 152 transit trips that make a total of 384 station stops. input folder contains all the supply-side/network input files prepared from the test network. More information about network input file standards can be found in the GTFS-Plus Data Standards Repository.
Two versions of sample demand have been prepared:
demand_regcontains regular demand that consists only of a transit trip list. There are no multiple user classes and all trips use a single set of path weights (pathweight_ft.txt). Demand starts at 3:15 PM and ends at 5:15 PM.One trip occurs every 10 seconds. More information is available in documentation.demand_twopathsrepresents demand for two user classes that use different sets of path weights. Household and person attribute files are present in addition to the trip list to model user heterogeneity and multiple user classes.
Similar to network data standards, there also exists a Demand Data Standards Repository.
There are a total of six test runs in \scripts\runAllTests.bat. Type of assignment, capacity constraint, and number of iterations are varied in addition to the demand.
| Sno | Demand | Assignment Type | Iterations | Capacity Constraint |
|---|---|---|---|---|
| 1 | Multi-class | Deterministic | 2 | On |
| 2 | Multi-class | Stochastic | 1 | Off |
| 3 | Multi-class | Stochastic | 2 | On |
| 4 | Regular | Deterministic | 2 | On |
| 5 | Regular | Stochastic | 1 | Off |
| 6 | Regular | Stochastic | 2 | On |
Type of Assignment:
- "Deterministic" indicates use of a deterministic trip-based shortest path search algorithm
- "Stochastic" indicates use of a stochastic hyperpath-finding algorithm
-
Ramming, M. S. Network Knowledge and Route Choice. Ph.D. Thesis. Massachusetts Institute of Technology, Cambridge, Mass., 2002.
-
Hoogendoorn-Lanser, S., R. Nes, and P. Bovy. Path Size Modeling in Multinomial Route Choice Analysis. 27 In Transportation Research Record: Journal of the transportation Research Board, No 1921, 28 Transportation Research Board of the National Academies, Washington, D.C., 2005, pp. 27-34.
Major changes to fast-trips since the original FAST-TrIPs (https://github.com/MetropolitanTransportationCommission/FAST-TrIPs-1)
To be filled in further but including:
- Implemented overlap pathsize correction (8/2016)
- Add purpose segmentation to cost weighting (7/2016)
- Output pathsets in addition to chosen paths (4/2016)
- Update transit trip vehicle times based on boards, alights and vehicle-configured accleration, deceleration and dwell formulas (4/2016)
- Output performance measures (pathfinding and path enumeration times, number of stops processed) (3/2016)
- Stop order update to pathfinding: when a stop state is updated, mark other reachable stops for reprocessing (3/2016) details
- Support KNR and PNR access (11/2015)
- Read user-class based cost weighting (11/2015)
- Switch input format to GTFS-plus network (10/2015)
- Move path finding to C++ extension (9/2015)
- Parallelized path finding with multiprocessing (7/2015)
- Port original FAST-TrIPs codebase to python with debug tracing (5/2015)