Micro-simulation of Traffic

Learning Objectives

  • Describe microscopic vehicle-following, lane-changing, and route-choice rules and their limitations.
  • Build a connected network with observed geometry, controls, traffic demand, and multimodal inputs.
  • Choose defensible warm-up periods, random seeds, run lengths, and output measures.
  • Separate model error checking, calibration, and independent validation.
  • Compare scenarios with uncertainty, sensitivity tests, reproducible runs, and responsible interpretation.
Microscopic traffic simulation models individual movements to investigate congestion mechanisms and proposed operating changes. It is a calibrated analysis model, not an animation that automatically proves traffic outcomes.

Microscopic traffic simulation

Time-stepped or event-based representation of individual vehicles and, when supported, pedestrians and other modes whose motions depend on a network, demand, control, and behavioral rules.

Model the physical network first

Nodes, links, lanes, controls, and observed demand

A network requires correct link connectivity and direction, lane storage lengths, allowable movements, signal timing and offsets, bus stops, pedestrian crossings, turning restrictions, origins/destinations, and time-varying arrivals. Field observations should cover congested and noncongested periods, vehicle classes, queues, travel times, route proportions, and bottleneck discharge. Local Philippine conditions (jeepney and bus dwell, motorcycles, pedestrian conflicts, informal loading) must be represented only to the extent supported by observations and the selected software.

A short link can create upstream spillbackLinks are directional and their length matters; a queue at the downstream junction may reach and block the upstream junction.Time-varying inputStorage / queue interactionSignalsIllustrative coded topology

Vehicle behavior and space gaps

Car-following model

A behavioral model that updates a vehicle's acceleration or speed using the relative speed, spacing, and other modeled properties of a leader and follower.

Following is not a single universal law

Models include stimulus-response, psycho-physical (including Wiedemann families), safe-distance, and intelligent-driver formulations. Software implementations and parameter meanings differ. Drivers may respond to gap, relative speed, desired speed, reaction delay, and constraints; default parameters are starting points, not a local Philippine calibration. An implausible queue may result from network or demand errors rather than from following behavior.

Differentiate front-to-front spacing from clear gapLeader front is 90 m and follower front 60 m along a lane, and the 5 m leader length gives a 25 m bumper gap. Distances are synthetic.FollowerLeaderGap: leader rear − follower frontIllustrative clear gap

Clear spacing between two vehicles

One-dimensional, same-lane classroom geometry; positions measured in metres along a common direction.

s=xL−xF−ℓLs=x_L-x_F-\ell_L

Spacing symbols and caution

ss is bumper-to-bumper clear gap (m), xLx_L and xFx_F are leader and follower front positions (m), and ℓL\ell_L is leader length (m). This is geometry, not a calibrated safe-gap requirement; if follower speed is positive, time headway h=s/vFh=s/v_F (s) is a descriptive diagnostic, not a universal driver-response threshold.

Basic time-step and demand conversion

For a constant-speed classroom step, Δx=vΔt\Delta x=v\Delta t, with travel distance Δx\Delta x in metres, speed vv in m/s, and duration Δt\Delta t in seconds. A counted demand of NN vehicles in Δt\Delta t hours has average hourly arrival rate q=N/Δtq=N/\Delta t (veh/h). Actual simulation updates can account for acceleration and a time-varying arrival profile.

Lane-change decision

The modeled decision to seek and execute a lane transition because of route necessity, lane availability, or travel-performance preference, subject to the software's gap and safety rules.

Mandatory versus discretionary lane changes

A mandatory change may be required to reach a turning pocket or avoid a lane drop. A discretionary change may seek a faster lane. The model must represent legal movements, target-lane gaps, interactions with motorcycles and buses, and available weaving distance. A vehicle can desire a change but be unable to execute it because the target lane is blocked.

A desired lane change requires an available gapThe upper lane ends at the right; its vehicle's planned path bends into a target lane containing leader and follower vehicles.Lane endsTarget-lane gap / follower response

Warm-up, random seeds, and observation windows

Control the experiment, not just the animation

A warm-up period builds realistic queues and occupancies before collecting output; test whether the chosen warm-up achieves a stable condition for the question being studied. Record warm-up duration, demand loading, output observation duration, simulation time step, and seed-control method. Use multiple independent random seeds for stochastic variation, and preferably paired seeds for comparable alternatives where supported. The number of runs depends on variability and the decision's required confidence—not a universal count of 10 or 20.

Warm-up output should not silently enter the comparison windowSynthetic experiment timeline uses a 15-minute warm-up and 60-minute measurement window; the values are examples, not prescribed defaults.Warm-up 15 minMeasure 60 minStart15 min75 minSeed 1: compare output measuresSeed 2, 3, …: same horizon, alternate realizations

Estimated standard error of a replicated mean

Use when seed-level replications are approximately independent and a sample standard deviation is meaningful.

SE(Yˉ)=sYnSE(\bar{Y})=\frac{s_Y}{\sqrt{n}}

Reported seed uncertainty

sYs_Y is the sample standard deviation of seed-level outputs (same units as measure YY), nn is the number of replications, and SESE is the estimated standard error (same units as YY). A confidence interval also requires the chosen confidence level and appropriate sampling assumptions. Describe changes in peak queues, bottleneck timing, and distributions, not only a mean delay.

Paired-run difference summary

For paired run differences d1,…,dnd_1,\ldots,d_n, compute their arithmetic mean dˉ=(∑idi)/n\bar d=(\sum_i d_i)/n, their sample standard deviation sd=∑i(di−dˉ)2/(n−1)s_d=\sqrt{\sum_i(d_i-\bar d)^2/(n-1)} for n>1n>1, and the paired mean's estimated standard error sd/ns_d/\sqrt n. This describes stochastic variation for the replicated experiment, not model-input or future-demand uncertainty.

Calibration, validation, and forecasting

Match multiple observed phenomena

Error checking verifies links, signal programs, routing, units, origin/destination reconciliation, and impossible movements. Calibration tunes defensible parameters against designated observations such as travel-time distributions, volumes, queues, and bottleneck onset. Validation checks an independent dataset not used for parameter adjustment. Set objective-specific acceptance criteria before tuning; FHWA guidance is methodological, not an automatically applicable Philippine numeric threshold. Investigate failure mechanisms rather than changing parameters to produce a preferred outcome.

A single mean can hide the queue build-up errorSynthetic queues are 0, 4, 9, 15, 12, 6 vehicles observed and 0, 3, 7, 10, 10, 5 simulated; mismatch near the peak calls for investigation.Queue (veh)Observed —Model - -Successive time periods →

GEH hourly count diagnostic

A comparative count-fit statistic using observed and simulated volumes for the same hour; not an automatic acceptance rule.

GEH=2(C−M)2C+MGEH=\sqrt{\frac{2(C-M)^2}{C+M}}

Interpreting the GEH diagnostic

CC is the simulated hourly volume and MM is the observed hourly volume at the same count station. Both must use the same time window and units. This heuristic depends on the flow-count scale and does not establish a universal pass/fail threshold or replace other observed performance checks.

Mean absolute percentage error

Use only when all observed values are strictly positive; retain model bias direction separately.

MAPE=100m∑i=1m∣Ysim,i−Yobs,iYobs,i∣MAPE=\frac{100}{m}\sum_{i=1}^{m}\left|\frac{Y_{\mathrm{sim},i}-Y_{\mathrm{obs},i}}{Y_{\mathrm{obs},i}}\right|

Interpreting relative errors

mm is the number of matched periods; Yobs,iY_{\mathrm{obs},i} and Ysim,iY_{\mathrm{sim},i} are corresponding observed and simulated values in identical units. Individual absolute percentage errors use the same ratio without averaging. A single average can mask a severe peak queue or opposing signed biases.

Controlled calibration and alternatives workflow

  1. Define the decision, study network, representative operating conditions, output measures, and independent datasets.
  2. Build geometry, signals, demand, route choices, vehicle classes, and measured time profiles.
  3. Error-check the model before adjusting driver parameters.
  4. Calibrate against preselected observed measures and record all changes and seeds.
  5. Validate against independent data; if inadequate, investigate data/model causes and repeat relevant checks.
  6. Freeze an accepted base model, design no-build/build alternatives, select replicated seeds and common measurement windows.
  7. Test sensitivity, report seed variability and confidence, and disclose limitations and raw results.
Microsimulation Calibration and Validation

Build and error-check a microscopic traffic model, calibrate it to local driving behavior, validate it with independent observations, and gate forecasting on validation.

Calibration uses local real-world observations. Validation must use independent field data that was not used during calibration. The lesson gives no universal numerical match threshold, so the diagram does not invent one. A failed validation result does not authorize forecasting or imply a repair method.

Microsimulation Calibration and ValidationBuild and error-check a microscopic traffic model, calibrate it to local driving behavior, validate it with independent observations, and gate forecasting on validation.. Calibration uses local real-world observations. Validation must use independent field data that was not used during calibration. The lesson gives no universal numerical match threshold, so the diagram does not invent one. A failed validation result does not authorize forecasting or imply a repair method.. Start model development → Build network and enter traffic demand; Build network and enter traffic demand → Inspect animation for network or signal coding errors; Inspect animation for network or signal coding errors → Adjust behavior parameters using local observations; Adjust behavior parameters using local observations → Local behavior matched?; Local behavior matched? — Yes → Compare with independent field data; Local behavior matched? — No → Adjust behavior parameters using local observations; Compare with independent field data → Independent data matched?; Independent data matched? — Yes → Use the validated model to test alternatives; Independent data matched? — No → Hold forecasting; validation has not established readiness; Hold forecasting; validation has not established readiness → End model workflow; Use the validated model to test alternatives → Run multiple random seeds and average results; Run multiple random seeds and average results → End model workflow

Start model development → Build network and enter traffic demand; Build network and enter traffic demand → Inspect animation for network or signal coding errors; Inspect animation for network or signal coding errors → Adjust behavior parameters using local observations; Adjust behavior parameters using local observations → Local behavior matched?; Local behavior matched? — Yes → Compare with independent field data; Local behavior matched? — No → Adjust behavior parameters using local observations; Compare with independent field data → Independent data matched?; Independent data matched? — Yes → Use the validated model to test alternatives; Independent data matched? — No → Hold forecasting; validation has not established readiness; Hold forecasting; validation has not established readiness → End model workflow; Use the validated model to test alternatives → Run multiple random seeds and average results; Run multiple random seeds and average results → End model workflow

  • Start model development: terminator
  • Build network and enter traffic demand: process
  • Inspect animation for network or signal coding errors: process
  • Adjust behavior parameters using local observations: process
  • Local behavior matched?: decision
  • Compare with independent field data: process
  • Independent data matched?: decision
  • Hold forecasting; validation has not established readiness: process
  • Use the validated model to test alternatives: process
  • Run multiple random seeds and average results: process
  • End model workflow: terminator

Reproducibility and calibration are not the same as prediction certainty

A model may fit one day and fail under a different congestion regime. Report demand uncertainty, driver-behavior sensitivity, construction assumptions, transit/pedestrian representation, and model boundary effects. Do not label an alternative “better” when its apparent improvement is smaller than run-to-run uncertainty.
Key Takeaways
  • Vehicle trajectories depend on connected geometry, observed demand, controls, and calibrated behavior.
  • Lane changing includes route constraints and accepted gaps, not just lane-speed preferences.
  • Warm-up, seeds, replication, and consistent observation windows make comparisons interpretable.
  • Calibration uses tuning observations; validation uses independent evidence.
  • Scenario differences need uncertainty and sensitivity analysis, not just persuasive animation.

References