• Randomness
Randomness

Stochasticity is an important aspect of reproducing reality in a simulation scenario. There are multiple ways of adding stochasticity to a simulation. They are described below.

Random number generation (RNG)#

Sumo implements the Mersenne Twister algorithm for generating random numbers. This random number generator (RNG) is initialized with a seed value which defaults to the (arbitrary) value 23423. This setup makes all applications deterministic by default as the sequence of random numbers is fixed for a given seed. The seed may be changed using the option --seed <INT>. When using the option --random the seed will be chosen based on the current system time resulting in truly random behavior.

The simulation uses multiple RNG instances to decouple different simulation aspects

  • randomness when loading vehicles (typeDistribution, speed deviation,...)
  • probabilistic flows
  • vehicle driving dynamics
  • vehicle devices

The decoupling is done to ensure that loading vehicles does not affect simulation behavior of earlier vehicles. All RNGs use the same seed.

Route Distributions#

Vehicles can be added to the simulation with a fixed route (<vehicle>) or with an origin-destination pair (<trip>). A third alternative is to specify a set of routes (<routeDistribution>) and let the vehicle draw a random route from such a distribution. For details, see route distributions.

Vehicle Type Distributions#

A simple way of of modelling a heterogeneous vehicle fleet works by defining a <vTypeDistribution> and let each vehicle pick it's type randomly from this distribution. For details, see vehicle type distributions.

Speed distribution#

By default, vehicles in SUMO adhere to the maximum speed defined for the lane they are driving on (if the maximumSpeed of their vehicle type allows it). This behavior can be modified using the <vType>-attribute speedFactor which makes vehicles drive with that factor of the current speed limit. The attribute also allows the specification of the parameters of a normal distribution with optional cutoffs. The random value is selected once for each vehicle at the time of its creation. Using a speed deviation is the recommended way for getting a heterogenous mix of vehicle speeds. By default, a speed distribution with a standard deviation of 10% is active. For details, see speed distribution

Car-Following#

The default car-following model Krauss supports stochastic driving behavior through the vType-attribute sigma (default 0.5). When this value is non-zero drivers will randomly vary their speed based on the RNG described above. Other car-following models also use this attribute.

Departure times#

The departure times of all vehicles may be varied randomly by using the option --random-depart-offset <TIME>. When this option is used each vehicle receives a random offset to its departure time, equidistributed on [0, <TIME>].

Flows with a random number of vehicles#

The following features for random flows apply to duarouter and sumo

Binomial distribution#

By definining a <flow> with attributes end and probability (instead of vehsPerHour,number, or period), a vehicle will be emitted randomly with the given probability each second until the end time is reached. The number of vehicles inserted this way will be binomially distributed. When modeling such a flow on a multi-lane road it is recommended to define a <flow> for each individual lane because the insertion rate is limited to at most 1 vehicle per second.

Note

The effective flow may be higher at lower step-length because the discretization error is reduced (vehicles usually cannot be inserted in subsequent seconds due to safety constraints and insertion in every other second does not achieve maximum flow).

For low probability the distribution of inserted vehicles approximates a Poisson Distribution

Poisson distribution#

Since version 1.13.0 flow can also be defined with attribute end and period="exp(X)" where X is a positive value. This will cause the time gaps between vehicle insertions to follow an exponential distribution with rate parameter X. Effectively insertion an expected value of X vehicles per second. The number of vehicles inserted this way will follow the Poisson distribution.

Note

The effective insertion rate is limited by network capacity and other flow attributes such as departSpeed and departLane

Flows with a fixed number of vehicles#

The following 2 sections describe attributes for random flows that apply to duarouter and sumo. They are quite similar to flows with a random number of vehicles but substitute the number attribute for the end attribute.

Bernoulli Process#

By definining a <flow> with attributes number and probability (instead of vehsPerHour,number, or period), a vehicle will be emitted randomly with the given probability each second until the specified number is reached.

Poisson Process#

By definining a <flow> with attributes number and period="exp(X)" (instead of vehsPerHour,number, or period), vehicles will emitted with random time-gaps that follow an exponential distribution until the specified number is reached.

Router options#

The duarouter, dfrouter and jtrrouter applications support the option --randomize-flows.

When this option is used, each vehicle defined by a <flow>-element will be given a random departure time which is equidistributed within the time interval of the flow. (By default vehicles of a flow are spaced equally in time). The departure times computed this way also achieve a Poisson process

Departure and arrival attributes#

The <flow>, <trip> and <vehicle> elements support the value "random" for their attributes departLane, departPos, departSpeed and arrivalPos. The value will be chosen randomly on every insertion try (for the departure attributes) or whenever there is a need to revalidate the arrival value (i.e. after rerouting). The attribute departPosLat also supports the value "random". The lateral offset at departure will only affect simulation behavior when using the sublane model though it will be visible without this model too.

Lateral Variation#

When setting the lane change mode attribute lcSigma to a positive value, Vehicles will exhibit some random lateral drift.

Further sources of randomness#

Reproducibility#

Generally, all SUMO applicaitons and tools are expected to produce the same results when running the same version repeatedly with the same arguments and inputs. This also includes running on different platforms (Windows/Linux/Mac).

There are some situations that are know to violate this rule (either by design or due to technical reasons) and they are listed in the following.

Violated reproducibility by design#

Option --random is supported by many applications (sumo, duarouter, ...) and tools (randomTrips.py, ...). It randomizes the random seed and thereby gives a different random behavior each run.

Violated reproducibility for other reasons#

The following differences are either due to bugs or hard-to-solve problems in libraries used by SUMO.

Platform differences#

  • Differences in generated networks by platform: When importing network data with geo-coordinates, the transformation to x,y (Cartesian) coordinates are performed by the Proj. This library has differences from one version to the next and differnt platforms usually provide different library versions. The same problem may also manifest when using TraCI-functions for coordinate transformations.
  • Non-deterministic vehicle routing with options --device.rerouting.threads --weights.random-factor (Issue #10292)
  • Different Simulation behavior due to using the log function. Observed differences for the EIDM-Model (Issue #8921) and also for Simulations with the DriverState-device. Could in principe also affect the Wiedemann-Model and the ToC-device.