sumoITScontrol: Traffic Control Benchmarking

• sumoITScontrol is an open-source Python framework for controlling and benchmarking traffic management strategies in SUMO, ensuring reproducible experiments.
• The framework separates network definition, sensor mapping, and controller logic to support variance-aware, statistically robust simulation studies.
• It implements diverse control strategies—ramp metering (ALINEA, METALINE, HERO) and signal management (MaxPressure, ScootScats)—for both urban and freeway settings.

sumoITScontrol is an open-source Python framework for controlling and benchmarking traffic management strategies in the microscopic simulator SUMO via the TraCI interface. It provides a curated collection of implemented traffic controllers for both urban and freeway settings, including Max Pressure signal control, SCOOT/SCATS-inspired adaptive coordination, and ramp metering algorithms such as ALINEA, HERO, and METALINE. Its stated purpose is not only to supply reusable controller implementations, but also to improve reproducibility, comparability, and methodological transparency in intelligent traffic control experiments conducted in stochastic microscopic environments (Riehl et al., 25 Apr 2026).

1. Definition, scope, and research motivation

sumoITScontrol is positioned as a response to two recurring problems in the SUMO and ITS literature: the lack of standard baselines and weak evaluation methodology. The framework targets the first problem by providing well-documented, modular implementations of widely used traffic controllers, and the second by emphasizing multi-replication experiments, stochastic calibration, variance-aware reporting, and formal statistical testing (Riehl et al., 25 Apr 2026).

The framework is designed for two broad application domains. In urban networks, it addresses signalised intersections and arterial coordination. In freeway corridors, it addresses local and coordinated ramp metering. This dual scope is important because traffic control studies often treat these domains separately, whereas sumoITScontrol presents them within a unified software environment built around SUMO and TraCI (Riehl et al., 25 Apr 2026).

A central methodological claim of the framework is that microscopic simulations are intrinsically stochastic. Departures, lane changing, car-following, queue formation, and merge interactions vary across seeds, so single-run comparisons can confound controller effects with simulation noise. The framework therefore treats reproducible benchmarking as a first-class concern rather than an afterthought. This emphasis distinguishes it from project-specific controller implementations that are difficult to compare or replicate across studies (Riehl et al., 25 Apr 2026).

2. Software architecture and SUMO integration

The framework is implemented as a Python package that attaches to a running SUMO simulation through TraCI. A typical workflow consists of defining the network and demand in .net.xml, .rou.xml, and .sumocfg files, starting SUMO from Python, instantiating scenario abstractions and controllers, and then stepping the simulation while controllers read measurements and send commands back to SUMO. Measurements can be taken from loop detectors, lane area detectors, or lane-level state, and control actions include traffic signal phase updates and ramp metering rate adjustments (Riehl et al., 25 Apr 2026).

Its architecture separates network and demand definition, sensor mapping, and controller logic. This separation is reflected in the main abstractions exposed by the framework:

Abstraction	Role	Typical contents
RampMeter	Single ramp signal and sensors	tl_id, mainline sensors, queue sensors
RampMeterCoordinationGroup	Coordinated ramp set	Ordered list of RampMeter objects
Intersection	One signalised junction	tl_id, phases, links, state strings
IntersectionGroup	Coordinated signal group	Intersections, districts, corridor mappings

Controller classes are built on top of these abstractions. For ramp metering, the framework includes ALINEA, METALINE, and HERO. For signal control, it includes MaxPressure_Fix, MaxPressure_Flex, and ScootScats. Controllers expose parameters such as gains, thresholds, cycle lengths, and measurement periods, and implement a step() or update() style interface to be called either every simulation step or at specified control intervals (Riehl et al., 25 Apr 2026).

This design is explicitly extensible. New controllers can be introduced as Python classes that read TraCI-exposed measurements through the existing abstractions and apply new control actions back to SUMO. A plausible implication is that the framework is intended not merely as a repository of baselines, but also as an experimental substrate for new controller development, provided that the new implementations preserve the same reproducibility-oriented evaluation discipline.

3. Implemented controller families

sumoITScontrol implements two major controller families: freeway ramp metering and signalised intersection management (Riehl et al., 25 Apr 2026).

Ramp metering

For freeway control, the framework models time in discrete steps and uses measurements such as flow $f_{t,i}$, speed $v_{t,i}$, occupancy $c_{t,i}$, and ramp queue length $l_{t,i}$. Each metered on-ramp has a metering rate $r_i \in [r_{min}, r_{max}] \subset [0,1]$, interpreted as the fraction of green time in a control cycle (Riehl et al., 25 Apr 2026).

ALINEA is implemented as a local feedback controller that regulates each ramp independently to keep downstream occupancy near a target $c^\star$. Its standard law is

$r_t = r_{t-1} + K_P \bigl(c^{\star} - c_t\bigr).$

A PI variant is also supported: $$r_t = r_{t-1} + K_P \bigl(c^{\star} - c_t\bigr) + K_I \bigl(c_t - c_{t-1}\bigr).$$ The exposed parameters include target_occupancy, K_P, K_I, cycle_duration, measurement_period, and minimum and maximum metering rates (Riehl et al., 25 Apr 2026).

METALINE extends ALINEA to a coordinated multi-ramp, multi-input multi-output setting. With vectors of rates $\mathbf{r}_t$, occupancies $\mathbf{c}_t$, and target occupancies $v_{t,i}$0, the controller applies

$v_{t,i}$1

with gain matrices $v_{t,i}$2. Off-diagonal entries model inter-ramp coupling, and diagonal gain matrices reduce the controller to independent PI-ALINEA controllers (Riehl et al., 25 Apr 2026).

HERO combines local ALINEA control with queue management. Under normal conditions each ramp runs ALINEA. When a queue exceeds an activation threshold $v_{t,i}$3, that ramp becomes a master and upstream ramps become slaves. Slave metering is then adjusted using a queue-based target discharge calculation: $v_{t,i}$4 with hysteresis defined by a release threshold $v_{t,i}$5 (Riehl et al., 25 Apr 2026).

The framework deliberately excludes systems such as SWARM and ZONE because they require extensive network-wide modelling, detector infrastructure, and archived ITS data, and are described as less transparent and harder to calibrate and reproduce across networks (Riehl et al., 25 Apr 2026).

Signalised intersection management

For urban signal control, the framework represents a network as intersections $v_{t,i}$6, incoming and outgoing links $v_{t,i}$7 and $v_{t,i}$8, and movement phases $v_{t,i}$9. Green times satisfy

$c_{t,i}$0

and lost time between phases is denoted $c_{t,i}$1. Phase pressure is defined as

$c_{t,i}$2

where $c_{t,i}$3 is a function of the queue on link $c_{t,i}$4 (Riehl et al., 25 Apr 2026).

MaxPressure_Fix uses a fixed cyclic phase order and fixed cycle duration $c_{t,i}$5, but adaptively reallocates green splits at cycle boundaries according to measured phase pressure. Average pressure over a cycle is

$c_{t,i}$6

and the effective green time is

$c_{t,i}$7

If total pressure is positive, the initial split allocation is

$c_{t,i}$8

Bounds are then enforced and residual time is redistributed until the green times sum to $c_{t,i}$9 (Riehl et al., 25 Apr 2026).

MaxPressure_Flex is acyclic and event-driven. After the current phase has served at least its minimum green time, the controller checks whether

$l_{t,i}$0

If so, it switches to a phase with maximal pressure; otherwise it keeps the current phase and re-evaluates after an adaptive interval $l_{t,i}$1. Maximum green enforces a phase change to avoid starvation. The implementation is described as a finite state machine with states Start, Check_Pressures, Wait, Next_Phase, and Transition (Riehl et al., 25 Apr 2026).

ScootScats is a group-based adaptive signal controller inspired by SCOOT and SCATS. It coordinates a shared cycle length across a group of intersections, adapts local green splits based on degree of saturation, and adjusts offsets to support progression. Cycle length is updated from the global maximum degree of saturation $l_{t,i}$2 using thresholds $l_{t,i}$3, $l_{t,i}$4, and gain $l_{t,i}$5, while green split adjustment uses the most saturated lane and redistributes remaining effective green. Offset optimisation uses district-level congestion and travel time estimates

$l_{t,i}$6

with sequential offset adaptation along the critical corridor (Riehl et al., 25 Apr 2026).

The framework explicitly does not re-implement fixed-time or Webster-type controllers, simple actuated controllers, MPC, dynamic programming, or reinforcement learning controllers. The stated rationale is to focus on transparent, classical feedback controllers that can serve as robust baselines (Riehl et al., 25 Apr 2026).

4. Evaluation methodology and statistical discipline

A defining feature of sumoITScontrol is its insistence that controller evaluation in SUMO must be variance-aware. The framework recommends at least 10–20 repetitions per configuration and a paired experimental design in which different controllers are tested with the same set of random seeds. The purpose of the paired design is to reduce noise in performance differences and increase statistical power (Riehl et al., 25 Apr 2026).

The paper recommends reporting not only mean performance but also standard deviation, 95% confidence intervals, and graphical distributions such as boxplots or histograms. It further recommends formal significance testing: paired t-tests for approximately normal paired data, Wilcoxon signed-rank tests when normality is doubtful, Mann–Whitney U tests for unpaired comparisons, and ANOVA or non-parametric equivalents for multiple-controller studies (Riehl et al., 25 Apr 2026).

An illustrative example is the ALINEA tuning experiment comparing $l_{t,i}$7 and $l_{t,i}$8. The reported mean objective values are $l_{t,i}$9 and $r_i \in [r_{min}, r_{max}] \subset [0,1]$0, with standard deviations $r_i \in [r_{min}, r_{max}] \subset [0,1]$1 and $r_i \in [r_{min}, r_{max}] \subset [0,1]$2, and $r_i \in [r_{min}, r_{max}] \subset [0,1]$3. Using the pooled standard deviation

$r_i \in [r_{min}, r_{max}] \subset [0,1]$4

the resulting test statistic is approximately

$r_i \in [r_{min}, r_{max}] \subset [0,1]$5

with two-sided $r_i \in [r_{min}, r_{max}] \subset [0,1]$6, which is not significant at $r_i \in [r_{min}, r_{max}] \subset [0,1]$7 (Riehl et al., 25 Apr 2026).

This example functions as a methodological warning. Apparent improvements in mean performance can be indistinguishable from stochastic variation. A common misconception in SUMO-based controller evaluation is that a controller ranking established on one or a few runs is informative; the framework argues directly against that assumption and treats replication and hypothesis testing as necessary components of credible benchmarking (Riehl et al., 25 Apr 2026).

5. Case studies and empirical demonstrations

The framework is demonstrated on both freeway and urban scenarios. In the freeway case study, the network is a 4.1 km two-lane motorway with three metered on-ramps of approximately 200 m, simulated for 4,200 s with a 0.5 s time step. The demand model uses heterogeneous drivers with at least five aggressiveness levels per vehicle type and origin, and vehicle insertion is performed through probabilistic arrivals using probability, with recommended insertion settings including departSpeed="max", departLane="free", departPos="last", and insertionChecks="none" (Riehl et al., 25 Apr 2026).

This freeway study evaluates ALINEA, METALINE, and HERO. The reported findings are that ALINEA performance depends non-monotonically on $r_i \in [r_{min}, r_{max}] \subset [0,1]$8, that METALINE is sensitive to gain tuning and can stabilize one ramp while creating oscillations or large queues elsewhere, and that HERO can improve conditions at heavily congested ramps by coordinating upstream ramps through master/slave logic (Riehl et al., 25 Apr 2026).

In the urban case study, the network is a real arterial in Esslingen am Neckar, Germany, with 7 signalised intersections and approximately 96 signals, of which 5 intersections are controlled within the reported setup. The simulation runs from 09:00 to 24:00 with a 0.25 s time step. Demand is calibrated with loop detector data and municipal administration information, and the scenario includes 26,714 single vehicle trips, 11 bus lines, 26 stops, and a fleet composition based on German statistics: 81% passenger cars, 8.2% motorcycles, 4.6% vans, and 6.2% trucks, with 2% of passenger cars being EVs (Riehl et al., 25 Apr 2026).

The urban demonstrations compare MaxPressure_Fix, MaxPressure_Flex, and ScootScats. The reported signal timing patterns differ systematically: MaxPressure_Fix maintains a stable cyclic sequence with variable green splits; MaxPressure_Flex varies both sequence and duration over time; and ScootScats adapts cycle length, green splits, and potentially offsets, although in the example offset optimisation is not triggered because congestion differences are small (Riehl et al., 25 Apr 2026).

These examples are not presented as definitive controller rankings. Their significance lies in showing that the same framework can host different controller philosophies under the same calibrated stochastic environment, making cross-method comparison more reproducible and methodologically explicit.

6. Relation to prior SUMO ITS control research, limitations, and future directions

sumoITScontrol is best understood within a broader lineage of SUMO-based ITS control research. Earlier work on a semi-decentralized urban traffic strategy based on TUC used SUMO and TraCI to combine centralized signal optimization with a contention time window and local priority rules approximating V2V or V2I behavior (Farhi et al., 2015). Another line of work coupled JADE, TraSMAPI, and SUMO to implement traffic lights as multi-agent systems using Q-learning, demonstrating a tool-chain for decentralized adaptive control in microscopic simulation (Azevedo et al., 2016). sumoITScontrol differs from these efforts in emphasis: rather than presenting one new controller architecture, it organizes a collection of classical controllers and couples them to explicit benchmarking guidance (Riehl et al., 25 Apr 2026).

This suggests a specific conceptual role for the framework. It is not a replacement for controller-specific research such as semi-decentralized TUC variants or multi-agent reinforcement learning. Rather, it provides a common baseline layer against which such approaches can be tested more transparently. A plausible implication is that its main scholarly value lies as much in standardization of experimental practice as in software implementation.

The paper also states clear limitations. The current controller set is not exhaustive and does not yet include variable speed limits, perimeter control, learning-based strategies, or fully automated benchmarking workflows. The reported case studies cover only one freeway corridor and one urban arterial. The current setup does not explicitly model sensor noise, communication failures or delays, or mixed connected-vehicle penetration (Riehl et al., 25 Apr 2026).

Planned future directions include extension to variable speed limit and perimeter control strategies, integration of learning-based algorithms on top of the baseline controllers, development of automated benchmarking workflows, and inclusion of sensor noise and CAV/EV aspects for robustness testing (Riehl et al., 25 Apr 2026). In that sense, sumoITScontrol occupies a transitional position in the SUMO research ecosystem: it formalizes a baseline repertoire and an evaluation standard, while leaving open the addition of richer controller classes and more realistic experimental conditions.