SIGMA: Symmetry-aware, Intelligent, Geometric, Multi-faced Adaptive Traffic Control for Reliable Intelligent Transportation Systems

Pratham Payra; Jagadish; Tanmay Sen; Tanujit Chakraborty

cs.LGAug 18, 2026

SIGMA: Symmetry-aware, Intelligent, Geometric, Multi-faced Adaptive Traffic Control for Reliable Intelligent Transportation Systems

Authors: Pratham Payra, Jagadish, Tanmay Sen, Tanujit Chakraborty

Organizations: SQC & OR, Indian Statistical Institute · Sorbonne Center for Artificial Intelligence, Sorbonne University Abu Dhabi

Abstract

Traffic signal control is a challenging sequential decision-making problem requiring a balance between throughput, vehicle delay fairness, signal stability, predictable switching, and emergency vehicle prioritisation. Existing reinforcement learning (RL) approaches often rely on predefined objectives and offer limited adaptability to changing operational priorities. They may also generalise poorly across geometrically equivalent intersections. To address these limitations, SIGMA (Symmetry-aware, Intelligent, Geometric, Multi-faced Adaptive traffic control) is proposed as a multi-faced RL framework augmented with a large language model (LLM) for dynamic objective adaptation and orientation-invariant policy learning. Natural language emergency instructions are translated into priority vectors that guide a multi-faced actor-critic controller. Symmetry-aware rotational augmentation supports transfer across homogeneous four-way intersections, while an offline-to-online learning strategy enables stable initialisation and adaptation to evolving traffic conditions. Structural properties concerning emergency service levels, graceful degradation under LLM failure, and sensitivity to traffic demand perturbations are established, alongside bootstrap-based statistical validation. Experiments in SUMO across four urban intersections modelled on Kolkata traffic scenarios compare SIGMA with fixed-time, actuated, and Deep Q-Network controllers. Reductions in average waiting times, emergency waiting times, and queue lengths are observed, alongside improved traffic throughput. Empirical reliability and ablation analyses further assess robustness to component failures and geometric orientation. SIGMA integrates language-guided objective adaptation, multi-faced policy learning, and statistical reliability assessment into an adaptive framework for intelligent traffic control.

Figures & tables

Figure 1: Illustration of incoming and outgoing queue measurements at a four-way intersection. The incoming $l_{i}^{\mathrm{in}}$ and outgoing $l_{i}^{\text{out}}$ queue lengths, $i\in\{E,N,W,S\}$ , are measured within the corresponding approach-specific detection zones and used to construct the movement-level traffic-pressure representation.

Figure 2: Signal phases are organized into three pattern groups. Type 1 phases ( $G_{1}$ ) serve a single incoming approach. Type 2 phases ( $G_{2}$ ) allow straight-through and left-turn movements from opposite approaches. Type 3 phases ( $G_{3}$ ) allow right-turn and U-turn movements from opposite approaches. In total, eight admissible phases are defined.

Figure 3: Overview of the SIGMA architecture, including the offline pretraining and online execution phases.

Figure 4: Examples of different types of action smoothness.

Objective	Objective function	Actor loss	Reward
Markovian consistency	$\displaystyle\mathcal{U}_{\mathrm{M}}^{(t)}=\left\\|\boldsymbol{\pi}^{(t)\top}\mathcal{Q}-\boldsymbol{\pi}^{(t+1)\top}\right\\|_{2}^{2}$	$\displaystyle\mathcal{L}_{\mathrm{M}}=\frac{1}{T-1}\sum_{t=1}^{T-1}\mathcal{U}_{\mathrm{M}}^{(t)}$	$\displaystyle r_{\mathrm{M}}^{(t)}=-\mathcal{U}_{\mathrm{M}}^{(t)}$
Action smoothness	$\displaystyle\mathcal{U}_{\mathrm{S}}^{(t)}=\left\\|\left(\boldsymbol{\pi}^{(t+1)}-\boldsymbol{\pi}^{(t)}\right)^{\top}A_{\mathrm{TR}}\right\\|_{2}^{2}$	$\displaystyle\mathcal{L}_{\mathrm{S}}=\frac{1}{T-1}\sum_{t=1}^{T-1}\mathcal{U}_{\mathrm{S}}^{(t)}$	$\displaystyle r_{\mathrm{S}}^{(t)}=-\mathcal{U}_{\mathrm{S}}^{(t)}$
Queue-pressure reduction	$\displaystyle\mathcal{U}_{\mathrm{Q}}^{(t)}=\left(\frac{1}{4}\boldsymbol{\pi}^{(t)\top}A_{\mathrm{TR}}\mathbf{M}_{q}^{t}-\max_{a\in\mathcal{A}}\frac{1}{4}{a}^{\top}A_{\mathrm{TR}}\mathbf{M}_{q}^{t}\right)^{2}$	$\displaystyle\mathcal{L}_{\mathrm{Q}}=\frac{1}{T}\sum_{t=1}^{T}\mathcal{U}_{\mathrm{Q}}^{(t)}$	$\displaystyle r_{\mathrm{Q}}^{(t)}=-\mathcal{U}_{\mathrm{Q}}^{(t)}$
Waiting-time fairness	$\displaystyle\mathcal{U}_{\mathrm{W}}^{(t)}=\left(\frac{1}{4}\boldsymbol{\pi}^{(t)\top}A_{\mathrm{TR}}\mathbf{M}_{w}^{t}-\max_{a\in\mathcal{A}}\frac{1}{4}{a}^{\top}A_{\mathrm{TR}}\mathbf{M}_{w}^{t}\right)^{2}$	$\displaystyle\mathcal{L}_{\mathrm{W}}=\frac{1}{T}\sum_{t=1}^{T}\mathcal{U}_{\mathrm{W}}^{(t)}$	$\displaystyle r_{\mathrm{W}}^{(t)}=-\mathcal{U}_{\mathrm{W}}^{(t)}$
Emergency alignment	$\displaystyle\mathcal{U}_{\mathrm{E}}^{(t)}=\left\\|\boldsymbol{\pi}^{(t)\top}A_{\mathrm{TR}}-\boldsymbol{\theta}^{t\top}\right\\|_{2}^{2}$	$\displaystyle\mathcal{L}_{\mathrm{E}}=\frac{1}{T}\sum_{t=1}^{T}\mathcal{U}_{\mathrm{E}}^{(t)}$	$\displaystyle r_{\mathrm{E}}^{(t)}=-\mathcal{U}_{\mathrm{E}}^{(t)}$

Table 1: Utility functions and their loss and reward components. The actor outputs the distribution

\pi^{(t)}

, and the action is

a^{t}=\operatorname{softmax}(\pi^{(t)})

Metric	Significance	Formulae
AEWT Δt	Emergency vehicle waiting time averaged over each evaluation interval $\Delta t$ . This metric is critical for assessing how well emergency vehicles are prioritised.	$\displaystyle\text{AEWT}_{\Delta t}=\frac{1}{k}\sum_{i=1}^{k}\bigl(t_{\text{em(dep),i}}-t_{\text{em(arr),i}}\bigr)$
AEWT	Overall emergency waiting time averaged over the full simulation period $T$ . This is the primary indicator of emergency response performance.	$\displaystyle\text{AEWT}=\frac{1}{T/\Delta t}\sum\text{AEWT}_{\Delta t}$
AWT Δt	Waiting time of all vehicles (regular and emergency) averaged over each interval $\Delta t$ . This reflects general user delay.	$\displaystyle\text{AWT}_{\Delta t}=\frac{1}{k}\sum_{i=1}^{k}\bigl(t_{\text{dep},i}-t_{\text{arr},i}\bigr)$
AWT	Overall waiting time of all vehicles averaged over the entire evaluation period. This is a key measure of average traffic efficiency.	$\displaystyle\text{AWT}=\frac{1}{T/\Delta t}\sum\text{AWT}_{\Delta t}$
AMWT Δt	Per-direction maximum waiting times averaged over the four approaches within interval $\Delta t$ . This captures the worst-case delay on any approach.	$\displaystyle\text{AMWT}_{\Delta t}=\frac{1}{4}\sum_{j=1}^{4}\left(\max_{i}^{k_{j}}\bigl(t_{\text{dep},i,j}-t_{\text{arr},i,j}\bigr)\right)$
PC t	Magnitude of phase changes between consecutive decision steps. Lower values indicate smoother transitions, which reduce driver confusion and jerkiness.	$\displaystyle\text{PC}_{t}=\frac{\\|a_{t}\cdot A_{\text{TR}}-a_{t+1}\cdot A_{\text{TR}}\\|_{2}^{2}}{8}$

Table 2: Evaluation Metrics for Traffic Signal Control

Method	AMWT (s) $\downarrow$	AWT (s) $\downarrow$	AEWT (s) $\downarrow$	AQL $\downarrow$	APC $\downarrow$	TC $\uparrow$	ATP $\uparrow$
Fixed Time Control	$231.47\pm 27.8$	$108.23\pm 13.0$	$91.15\pm 10.9$	$18.0\pm 1.8$	$0.50\pm 0.00$	$\mathbf{1.00\pm 0.00}$	$4.23\pm 0.48$
Actuated Control	$187.32\pm 22.5$	$99.61\pm 11.9$	$87.44\pm 10.5$	$16.0\pm 1.6$	$\mathbf{0.17\pm 0.02}$	$0.21\pm 0.03$	$6.31\pm 0.72$
DQN	$\mathbf{121.78\pm 14.6}$	$86.29\pm 10.4$	$29.55\pm 3.5$	$13.0\pm 1.3$	$0.38\pm 0.05$	$0.32\pm 0.04$	$8.07\pm 0.10$
SIGMA (Proposed)	$133.27\pm 16.3$	$\mathbf{77.54\pm 9.3}$	$\mathbf{23.47\pm 2.8}$	$\mathbf{11.38\pm 1.4}$	$0.47\pm 0.05$	$0.39\pm 0.05$	$\mathbf{9.26}\pm\mathbf{1.08}$

Table 3: Performance comparison of traffic signal control methods over 1000 episodes (After convergence of DQN and SIGMA from 2001-3000 th episode) (mean

\pm

standard deviation)

Panel of Figure 5 — Figure 5: Performance comparison across six metrics. Proposed method (red) shows improved emergency handling with minimal impact on general traffic. (a)–(d) are evaluated after 2,000 episodes, (e) over Episodes 2001–2050, and (f) shows convergence over the first 2,000 episodes comparing SIGMA and DQN.

Panel of Figure 6 — Figure 6: OSM-based representations of the four Kolkata intersections used for SUMO validation.

Figure 7: Comparative SUMO performance of the evaluated signal-control methods across the four Kolkata intersections.

Case	Emergency	Waiting Time	Queue Length	Markovian	Smoothness
	$(\lambda_{E},\alpha_{E})$	$(\lambda_{W},\alpha_{W})$	$(\lambda_{Q},\alpha_{Q})$	$(\lambda_{M},\alpha_{M})$	$(\lambda_{S},\alpha_{S})$
Case 1: Full Model	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$
All components enabled
Case 2: No Emergency Priority	$\times$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$
Emergency objective disabled
Case 3: No Emergency + No MWT	$\times$	$\times$	$\checkmark$	$\checkmark$	$\checkmark$

Table 4: Configurations considered in the objective-component ablation study. A check mark (

\checkmark

) indicates that the corresponding objective is enabled, whereas a cross (

\times

) indicates that it is excluded from the optimisation. Case 5 represents a uniform-weight configuration.

Case	AEWT (s) $\downarrow$	AMWT (s) $\downarrow$	AWT (s) $\downarrow$	AQL $\downarrow$	TC $\uparrow$	ATP $\uparrow$
Case 1: Full Model	$\mathbf{23.47\pm 2.8}$	$133.27\pm 16.3$	$77.54\pm 9.3$	$11.38\pm 1.4$	$0.39\pm 0.05$	$9.26\pm 1.08$
All components
Case 2: No Emergency	$72.37\pm 1.9$	$\mathbf{131.17\pm 15.9}$	$75.13\pm 8.1$	$10.97\pm 1.3$	$0.41\pm 0.07$	$9.31\pm 1.06$
$(\lambda_{E},\alpha_{E})$ disabled
Case 3: No Emergency + No MWT	$71.21\pm 1.7$	$177.27\pm 21.23$	$\mathbf{70.21\pm 6.9}$	$\mathbf{9.03\pm 1.1}$	$0.43\pm 0.04$	$\mathbf{10.18\pm 0.93}$
$(\lambda_{E},\alpha_{E})$ and $(\lambda_{W},\alpha_{W})$ disabled

Table 5: Performance of SIGMA under alternative objective-component configurations. Results are reported as mean

\pm

standard deviation over 1000 simulation episodes. Arrows indicate the preferred direction of each metric.

Figure 8: Training convergence of the RA and NRA configurations.

Training Configuration	AEWT (s) $\downarrow$	AMWT (s) $\downarrow$	AWT (s) $\downarrow$	AQL $\downarrow$
RA (Rotation-Augmented)	$\mathbf{23.47\pm 2.8}$	$\mathbf{133.27\pm 16.3}$	$77.54\pm 9.3$	$\mathbf{11.38\pm 1.4}$
Proposed training procedure with $C_{4}$ rotational augmentation
NRA (No Rotation Augmentation)	$24.18\pm 4.9$	$134.16\pm 21.26$	$\mathbf{76.97\pm 14.86}$	$12.21\pm 2.4$
Original orientation only

Table 6: Ablation study of rotation augmentation under the assumed fourfold directional symmetry. Results are reported as mean

\pm

standard deviation over the evaluation episodes.

Configuration	AEWT (s) $\downarrow$	AWT (s) $\downarrow$	AMWT (s) $\downarrow$	TC $\uparrow$	ATP $\uparrow$
C1: Balanced	$72.54$	$\mathbf{71.94}$	$176.17$	$0.42$	$10.03$
C2: Emergency	$\mathbf{18.23}$	$79.21$	$135.41$	$0.37$	$9.12$
C3: Queue	$74.39$	$73.16$	$142.37$	$0.34$	$\mathbf{10.41}$
C4: Fairness	$71.29$	$74.19$	$\mathbf{117.29}$	$0.36$	$9.53$
C5: Stability	$82.17$	$82.56$	$148.71$	$\mathbf{0.48}$	$8.47$
C6: E+Queue	$24.12$	$74.02$	$127.83$	$0.38$	$9.87$

Table 7: Performance summary under alternative objective-weight configurations. Boldface indicates the best value for each metric.

Figure 9: Accuracy-latency comparison of the evaluated language models for emergency-instruction interpretation.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

\mathcal{L}_{\text{actor}}=-\frac{1}{|\mathcal{B}|}\sum_{(s,a)\in\mathcal{B}}a^{\top}\log\pi_{\psi}(\cdot\mid s)+\lambda^{\top}\mathcal{L}_{R}

Appendix

Algorithm 1 Actor Pretraining

t_{1}\leftarrow\min\!\left(t_{0}+G_{\max},\inf\left\{t\geq t_{0}+G_{\min}:\max_{i}t_{l}^{(i)}(t)>G_{\mathrm{gap}}\right\}\right).

Appendix

Algorithm 5 Actuated Control

Pattern Group	Rotation Group	Size	Condition	Description
Single-Pivot Dominant	$G_{1}=\{A_{1}^{E},A_{1}^{N},A_{1}^{W},A_{1}^{S}\}$ Cyclic under $\rho_{a}$	4	Pivot $\mathbf{ic}\in\mathcal{IC}$	All movements from a single incoming direction.
Complementary Straight and Left-Turn	$G_{2}=\{A_{2}^{N-S},A_{2}^{E-W}\}$ Complements under $\rho_{a}$	2	$\mathbf{ic},\mathbf{og}$ opposite and opposite and left directions	Straight and left-turn movements for perpendicular pairs.
Complementary Right and U-Turn	$G_{3}=\{A_{3}^{N-S},A_{3}^{E-W}\}$ Complements under $\rho_{a}$	2	$\mathbf{ic},\mathbf{og}$ same and right directions	Right-turn and U-turn movements for perpendicular pairs.

Appendix

Table 8: Fundamental pattern groups for traffic signal actions under

C_{4}

symmetry.

$X\in\mathbb{R}^{16}$		$Y\in\mathbb{R}^{4}$		Direction Mapping
Rotation	Expression	Rotation	Expression
$0^{\circ}$	$X^{(0)}=I_{16}X$	$0^{\circ}$	$Y^{(0)}=I_{4}Y$	$E,N,W,S\rightarrow E,N,W,S$
$90^{\circ}$	$X^{(90)}=\rho X$	$90^{\circ}$	$Y^{(90)}=\rho^{\prime}Y$	$E,N,W,S\rightarrow N,W,S,E$
$180^{\circ}$	$X^{(180)}=\rho^{2}X$	$180^{\circ}$	$Y^{(180)}=\rho^{\prime 2}Y$	$E,N,W,S\rightarrow W,S,E,N$
$270^{\circ}$	$X^{(270)}=\rho^{3}X$	$270^{\circ}$	$Y^{(270)}=\rho^{\prime 3}Y$	$E,N,W,S\rightarrow S,E,N,W$

Appendix

Table 9: Rotation operators for

X\in\mathbb{R}^{16}

and

Y\in\mathbb{R}^{4}

under

C_{4}

symmetry.

\widehat{SL}(\tau)=\frac{1}{n}\sum_{i=1}^{n}\mathbb{I}\!\left(T_{em,i}<\tau\right)

Appendix

Algorithm 9 Percentile Bootstrap Confidence Interval for Emergency Service Level

Method	$\widehat{SL}(60)$	95% CI	CI Width
SIGMA (Full)	$0.992$	$[0.984,0.998]$	$0.014$
SIGMA-QW	$0.847$	$[0.812,0.881]$	$0.069$
DQN	$0.712$	$[0.624,0.794]$	$0.170$
Fixed-Time	$0.341$	$[0.258,0.424]$	$0.166$

Appendix

Table 10: Bootstrap Estimates of Emergency Service Level at a 60-s Threshold

$p_{\mathrm{fail}}$	MPD (AEWT)	MPD (AWT)	MPD (ATP)
$0.0$	$23.5$	$77.5$	$9.26$
$0.1$	$28.4$	$76.8$	$9.98$
$0.5$	$47.9$	$76.2$	$10.15$
$1.0$	$72.4$	$74.9$	$10.31$

Appendix

Table 11: Performance Under Simulated LLM Failure

Method	$\mathbf{S_{0.1}}$	$\mathbf{S_{0.2}}$
SIGMA	$0.18$	$0.35$
DQN	$0.41$	$0.82$
Fixed-Time	$0.05$	$0.11$

Appendix

Table 12: Sensitivity of Emergency Average Effective Waiting Time to Traffic-Demand Perturbations

$X\in\mathbb{R}^{k^{2}}$			$Y\in\mathbb{R}^{k}$
Rotation	Expression	Block Mapping	Rotation	Expression	Direction Mapping
$0$	$X^{(0)}=I_{k^{2}}X$	$0,1,\ldots,k-1\rightarrow 0,1,\ldots,k-1$	$0$	$Y^{(0)}=I_{k}Y$	$0,1,\ldots,k-1\rightarrow 0,1,\ldots,k-1$
$\frac{2\pi}{k}$	$X^{(1)}=\rho_{k}X$	$0,1,\ldots,k-1\rightarrow 1,2,\ldots,k-1,0$	$\frac{2\pi}{k}$	$Y^{(1)}=\rho^{\prime}_{k}Y$	$0,1,\ldots,k-1\rightarrow 1,2,\ldots,k-1,0$
$\frac{4\pi}{k}$	$X^{(2)}=(\rho_{k})^{2}X$	$0,1,\ldots,k-1\rightarrow 2,3,\ldots,k-1,0,1$	$\frac{4\pi}{k}$	$Y^{(2)}=(\rho^{\prime}_{k})^{2}Y$	$0,1,\ldots,k-1\rightarrow 2,3,\ldots,k-1,0,1$
$\vdots$	$\vdots$	$\vdots$	$\vdots$	$\vdots$	$\vdots$
$\frac{2\pi r}{k}$	$X^{(r)}=(\rho_{k})^{r}X$	cyclic block shift by $r$	$\frac{2\pi r}{k}$	$Y^{(r)}=(\rho^{\prime}_{k})^{r}Y$	cyclic shift by $r$

Appendix

Table 13: General rotation operators for

k^{2}

- and

k

-dimensional vectors under

C_{k}

symmetry.

Pattern Group	Rotation Behavior	Size	Condition	Description
Single-Pivot Dominant	$G_{1}^{(k)}=\{A_{1}^{(0)},A_{1}^{(1)},\ldots,A_{1}^{(k-1)}\}$ Cyclic under $\rho_{k}^{a}$	$k$	Pivot $\mathbf{ic}\in\mathcal{IC}_{k}$	All movements from a single incoming direction $i$ .
Complementary Pairs (if $k$ even)	$G_{2}^{(k)}=\{A_{2}^{(i)},A_{2}^{(i+k/2)}\}$ Complements under $\rho_{k}^{a}$	$k/2$	$\mathbf{ic},\mathbf{og}$ opposite directions	Straight and turning movements for antipodal pairs.
Additional Groups	$G_{m}^{(k)}$ (context-dependent)	Variable	Conflict graph dependent	Custom groups for non-conflicting movement sets.

Appendix

Table 14: Fundamental pattern groups for traffic signal actions under

C_{k}

symmetry.

Table 27

Step	TraCI API Call	Output Dim	Method	Description
1. Incoming queues	lane.getLastStepVehicleNumber() for $d\in D_{\text{in}}$	$\mathbb{R}^{4}$	All	Vehicle count per incoming approach
2. Outgoing queues	lane.getLastStepVehicleNumber() for $d\in D_{\text{out}}$	$\mathbb{R}^{4}$	All	Vehicle count per outgoing approach
3. Previous phase	trafficlight.getPhase(tls_id)	$\{0,1\}^{8}$	All	Current signal phase index encoded via $A_{TR}\to\{0,1\}^{16}$
4. Max waiting times	vehicle.getWaitingTime() per approach	$\mathbb{R}^{4}$	SIGMA-QW only	$\Gamma_{i}=\max_{veh\in\text{approach}_{i}}\text{waitingTime}(veh)$
5a. Build $\mathbf{M}_{q(\text{in})}$	repeat_each_4 ( $\mathbf{q}^{\text{in}}$ )	$\mathbb{R}^{16}$	SIGMA-QW only	$(l_{e}^{\text{in}},l_{e}^{\text{in}},l_{e}^{\text{in}},l_{e}^{\text{in}},l_{n}^{\text{in}},\dots,l_{s}^{\text{in}})^{\top}$
5b. Build $\mathbf{M}_{q(\text{out})}$	cycle_all_4 ( $\mathbf{q}^{\text{out}}$ )	$\mathbb{R}^{16}$	SIGMA-QW only	$(l_{e}^{\text{out}},l_{n}^{\text{out}},l_{w}^{\text{out}},l_{s}^{\text{out}},l_{e}^{\text{out}},\dots)^{\top}$

Appendix

Table 17: SUMO TraCI State Extraction for SIGMA-QW and Baseline Methods

Table 29

Config	$w_{E}$	$w_{W}$	$w_{Q}$	$w_{M}$	$w_{S}$	Focus
C1: Balanced	$0.20$	$0.20$	$0.20$	$0.20$	$0.20$	Equal emphasis
C2: Emergency	$0.50$	$0.15$	$0.15$	$0.10$	$0.10$	Prioritize emergency response
C3: Queue	$0.10$	$0.15$	$0.50$	$0.125$	$0.125$	Maximize throughput
C4: Fairness	$0.10$	$0.50$	$0.15$	$0.125$	$0.125$	Minimize max waiting time
C5: Stability	$0.10$	$0.15$	$0.15$	$0.30$	$0.30$	Smooth, predictable transitions
C6: E+Queue	$0.35$	$0.10$	$0.35$	$0.10$	$0.10$	Balance emergency & efficiency

Appendix

Table 20: Weight configurations for sensitivity analysis. Each row shows the weight vector

\mathbf{w}=(w_{E},w_{W},w_{Q},w_{M},w_{S})

with

\sum w_{k}=1

Configuration	AEWT (s)	AMWT (s)	AWT (s)	AQL	TC	ATP	Stability
C1: Balanced	$72.54$	$176.17$	$71.94$	$9.18$	$0.42$	$10.03$	Medium
C6: E+Queue	$24.12$	$127.83$	$71.23$	$10.26$	$0.38$	$9.87$	Medium
Markovian ( $\epsilon=0.01$ )	$82.34$	$152.18$	$83.91$	$12.87$	$\mathbf{0.48}$	$8.12$	Very High
Markovian ( $\epsilon=0.005$ )	$77.23$	$144.87$	$78.23$	$12.01$	$0.47$	$8.67$	Very High

Appendix

Table 22: Performance of the Markovian consistency-dominated regime (

\epsilon=0.01

Current $\rightarrow$ Next	$A_{1}^{E}$	$A_{2}^{E-W}$	$A_{3}^{E-W}$	$A_{1}^{W}$	$A_{1}^{N}$	$A_{2}^{N-S}$	$A_{3}^{N-S}$	$A_{1}^{S}$
$A_{1}^{E}$	$65.2\%$	$34.8\%$	$0.0\%$	$0.0\%$	$0.0\%$	$0.0\%$	$0.0\%$	$0.0\%$
$A_{2}^{E-W}$	$0.0\%$	$71.4\%$	$28.6\%$	$0.0\%$	$0.0\%$	$0.0\%$	$0.0\%$	$0.0\%$
$A_{3}^{E-W}$	$0.0\%$	$0.0\%$	$68.9\%$	$31.1\%$	$0.0\%$	$0.0\%$	$0.0\%$	$0.0\%$
$A_{1}^{W}$	$0.0\%$	$0.0\%$	$0.0\%$	$72.3\%$	$27.7\%$	$0.0\%$	$0.0\%$	$0.0\%$
$A_{1}^{N}$	$0.0\%$	$0.0\%$	$0.0\%$	$0.0\%$	$69.8\%$	$30.2\%$	$0.0\%$	$0.0\%$
$A_{2}^{N-S}$	$0.0\%$	$0.0\%$	$0.0\%$	$0.0\%$	$0.0\%$	$73.1\%$	$26.9\%$	$0.0\%$

Appendix

Table 23: Phase transition probabilities under the Markovian regime (

\epsilon=0.01

). Rows represent the current phase; columns represent the next phase (Figure 10 ).

$\epsilon$	AWT (s) $\downarrow$	AMWT (s) $\downarrow$	TC $\uparrow$
$0.001$	$85.34\pm 8.7$	$158.12\pm 18.3$	$\mathbf{0.49\pm 0.02}$
$0.005$	$82.89\pm 8.3$	$151.45\pm 17.1$	$0.48\pm 0.03$
$0.010$	$78.23\pm 8.1$	$144.87\pm 16.1$	$0.47\pm 0.03$
$0.025$	$75.12\pm 7.8$	$138.23\pm 15.4$	$0.45\pm 0.04$
$0.050$	$73.45\pm 7.5$	$133.89\pm 14.8$	$0.42\pm 0.04$

Appendix

Table 24: Sensitivity to

\epsilon

in the Markovian regime.

Figure 10: Sample transitions under Markovian regime

Traffic Signal Control

Traffic signal control studies how to set intersection phases and green times to reduce congestion while balancing safety, emissions, and priority needs.

See all 39 papers

Apr 7, 2026

CuraLight: Debate-Guided Data Curation for LLM-Centered Traffic Signal Control

Traffic signal control (TSC) is a core component of intelligent transportation systems (ITS), aiming to reduce congestion, emissions, and travel time. Recent approaches based on reinforcement learning (RL) and large language models (LLMs) have improved adaptivity, but still suffer from limited interpretability, insufficient interaction data, and weak generalization to heterogeneous intersections. This paper proposes CuraLight, an LLM-centered framework where an RL agent assists the fine-tuning of an LLM-based traffic signal controller. The RL agent explores traffic environments and generates high-quality interaction trajectories, which are converted into prompt-response pairs for imitation fine-tuning. A multi-LLM ensemble deliberation system further evaluates candidate signal timing actions through structured debate, providing preference-aware supervision signals for training. Experiments conducted in SUMO across heterogeneous real-world networks from Jinan, Hangzhou, and Yizhuang demonstrate that CuraLight consistently outperforms state-of-the-art baselines, reducing average travel time by 5.34 percent, average queue length by 5.14 percent, and average waiting time by 7.02 percent. The results highlight the effectiveness of combining RL-assisted exploration with deliberation-based data curation for scalable and interpretable traffic signal control.

Qing Guo, Xinhang Li, Junyu Chen +4

School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing, China · Beijing Big Data Center, Beijing, China

Traffic Signal Control Language Model-Based Control Training Data Curation LLM-Guided RL

Ask PDF

Mar 16, 2026

Evaluating the Robustness of Reinforcement Learning based Adaptive Traffic Signal Control

Reinforcement learning (RL) has attracted increasing interest for adaptive traffic signal control due to its model-free ability to learn control policies directly from interaction with the traffic environment. However, several challenges remain before RL-based signal control can be considered ready for field deployment. Many existing studies rely on simplified signal timing structures, robustness of trained models under varying traffic demand conditions remains insufficiently evaluated, and runtime efficiency continues to pose challenges when training RL algorithms in traffic microscopic simulation environments. This study formulates an RL-based signal control algorithm capable of representing a full eight-phase ring-barrier configuration consistent with field signal controllers. The algorithm is trained and evaluated under varying traffic demand conditions and benchmarked against state-of-the-practice actuated signal control (ASC). To assess robustness, experiments are conducted across multiple traffic volumes and origin-destination (O-D) demand patterns with varying levels of structural similarity. To improve training efficiency, a distributed asynchronous training architecture is implemented that enables parallel simulation across multiple computing nodes. Results from a case study intersection show that the proposed RL-based signal control significantly outperforms optimized ASC, reducing average delay by 11-32% across movements. A model trained on a single O-D pattern generalizes well to similar unseen demand patterns but degrades under substantially different demand conditions. In contrast, a model trained on diverse O-D patterns demonstrates strong robustness, consistently outperforming ASC even under highly dissimilar unseen demand scenarios.

Dickens Kwesiga, Angshuman Guin, Khaled Abdelghany +1

Research Engineer, School of Civil and Environmental Engineering, Georgia Institute of Technology, Atlanta, USA, 30332 · Principal Research Engineer, School of Civil and Environmental Engineering, Georgia Institute of Technology, Atlanta, USA, 30332 · Professor, Department of Civil and Environmental Engineering, Southern Methodist University, Dallas, USA, 75205 +1

Traffic Signal Control Reinforcement Learning Robust RL

Ask PDF

May 28, 2026

ReasonLight: A Multimodal Foundation Model-Enhanced Reinforcement Learning Framework for Zero-Shot Traffic Signal Control

Reinforcement learning (RL) has shown promise in traffic signal control (TSC). However, its reliance on predefined states limits responsiveness to observable open-world events that are absent from training data. IoT-enabled intersections provide heterogeneous observations from roadside sensors and cameras, creating opportunities to improve RL adaptability to such events. To this end, we propose ReasonLight, a multimodal foundation model-enhanced RL framework for zero-shot TSC. ReasonLight integrates three sources of information: structured traffic measurements, multi-view camera observations, and candidate phase decisions from a pre-trained RL controller. Given an RL-proposed phase, ReasonLight extracts visual semantics from multi-view images and aligns them with compact sensor-derived scene descriptions. This alignment enables a semantic-guided refinement module to either preserve or adjust the proposed action according to traffic rules and event semantics. To ensure operational reliability, refined actions are constrained by the set of available phases. Any invalid decision is rejected, and the system falls back to the original RL action. We evaluate ReasonLight on two types of rare events not seen during RL training: emergency vehicle priority and temporary traffic regulation. Experimental results show that ReasonLight achieves zero-shot adaptation without retraining. It reduces emergency vehicle waiting time by up to 88.7% compared with the RL-only backbone while preserving comparable routine traffic performance.

Aoyu Pang, Maonan Wang, Yuejiao Xie +3

School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, China · Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Hong Kong · Shanghai AI Laboratory, Shanghai, China +1

RL Control Intelligent Transportation Systems Traffic Signal Control Multimodal Foundation Models

Ask PDF

Abstract

Traffic Signal Control

Explore similar work

CuraLight: Debate-Guided Data Curation for LLM-Centered Traffic Signal Control

Evaluating the Robustness of Reinforcement Learning based Adaptive Traffic Signal Control

ReasonLight: A Multimodal Foundation Model-Enhanced Reinforcement Learning Framework for Zero-Shot Traffic Signal Control