⚡ electrical engineering

Active Inference for Adaptive Traffic Signal Control in Noisy Nonstationary IoT Environments

This paper proposes a traceable active inference controller for adaptive traffic signal management that outperforms both rule-based heuristics and deep reinforcement learning in noisy, nonstationary IoT environments by minimizing expected free energy to reduce idle times and CO2 emissions.

Original authors: Dénes Toth, George Ambroladze, Edwin Sundberg, Ali Beikmohammadi, Alfreds Lapkovskis

Published 2026-06-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dénes Toth, George Ambroladze, Edwin Sundberg, Ali Beikmohammadi, Alfreds Lapkovskis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a busy city intersection as a giant, chaotic kitchen where cars are the ingredients, traffic lights are the chefs, and the goal is to get everyone fed (moving) as quickly as possible without burning the food (causing pollution or accidents).

This paper introduces a new kind of "smart chef" for traffic lights called Active Inference (AIF). Here is how it works, how it was tested, and what it found, explained simply.

The Problem: The Kitchen is Foggy and Chaotic

In the real world, traffic sensors (like cameras or radar) aren't perfect. They get blocked by big trucks, confused by heavy rain, or just miss cars entirely. This is like a chef trying to cook while wearing foggy glasses and dealing with customers who keep changing their orders.

Old Rules: Traditional traffic lights follow a strict recipe (e.g., "Green for 30 seconds, then Red"). They can't adapt when the kitchen gets crazy.
The "Black Box" Learner: Newer AI systems (called Deep Reinforcement Learning or DQN) are like a sous-chef who has practiced cooking thousands of times. They are great at cooking if they have practiced in that exact same kitchen. But if the fog gets thicker or a customer suddenly orders something weird, they might get confused. Also, no one knows why they made a specific choice; they just do it.

The Solution: The "Curious Chef" (Active Inference)

The authors propose a new chef based on Active Inference. Instead of just memorizing a recipe or guessing based on past practice, this chef uses a special mental model to constantly ask two questions:

"What do I want to see?" (Pragmatic Value): "I want to see the traffic lines getting shorter and the air getting cleaner."
"What do I need to learn?" (Epistemic Value): "I'm not sure what's happening because of the fog. I should switch the light to get a better look at the other side of the street."

This chef tries to minimize a concept called "Expected Free Energy." Think of this as a "Confusion Score." The chef wants to make choices that lower their confusion and get them closer to their goal (smooth traffic).

The Best Part: Unlike the "Black Box" AI, this chef keeps a diary. You can look at the diary and see exactly why they switched the light: "I switched because I was worried I couldn't see the cars on the North side, so I turned the light to check."

The Experiment: The Simulation Kitchen

The researchers built a digital traffic simulation (a virtual kitchen) with a four-way intersection. They tested three chefs:

The Rule-Follower: Sticks to a fixed schedule.
The Trained AI (DQN): Learned by practicing for hours in the simulation.
The Curious Chef (AIF): Uses the "Confusion Score" method described above.

They ran the simulation through four levels of difficulty:

Level 1: Clear weather, normal traffic.
Level 2: Sensors get blocked by big trucks (Occlusion).
Level 3: It starts raining heavily (Weather).
Level 4: A car crashes, blocking the road, while it's raining and sensors are blocked (The "Perfect Storm").

The Results: Who Cooked Best?

1. In Calm Conditions (Levels 1 & 3):
The Trained AI (DQN) was slightly faster at moving cars. It had practiced for this specific scenario and knew the recipe by heart. The Curious Chef (AIF) was almost as good, but not quite the fastest.

2. In the "Perfect Storm" (Level 4):
This is where the Curious Chef (AIF) won big.

Idle Time: The Curious Chef kept cars waiting for 56,977 seconds total. The Trained AI kept them waiting for 71,741 seconds. That's a huge difference.
Pollution: The Curious Chef produced less CO2 emissions.
Why? When the sensors got messy and a crash happened, the Trained AI got confused because it hadn't practiced exactly that combination of disasters. The Curious Chef, however, was designed to handle uncertainty. It realized, "I don't know what's happening, so I need to gather more info and adapt," which kept traffic moving better.

3. The Trade-offs:
The Curious Chef wasn't perfect.

Bus Priority: It was slightly worse at letting buses through quickly compared to the others. Because it focused on the total number of cars to reduce overall confusion, it sometimes ignored a single bus waiting in the opposite lane.
Switching Lights: It changed the lights more often. While this helped it gather information, it might feel a bit "jittery" compared to the steady Rule-Follower.

The Bottom Line

This paper shows that while "practiced" AI is great for predictable situations, a Curious Chef that actively tries to understand uncertainty is much better when things go wrong (like bad weather or accidents).

The biggest win for the Curious Chef is transparency. If a traffic manager asks, "Why did you turn the light red?" the Curious Chef can explain its reasoning step-by-step. The Trained AI just says, "Because my brain told me to," which is hard to trust when things go wrong.

In short: When the traffic world is calm, a trained expert wins. When the world gets messy and unpredictable, a curious, self-aware system wins.

Technical Summary: Active Inference for Adaptive Traffic Signal Control in Noisy Nonstationary IoT Environments

1. Problem Statement

Urban traffic signal control faces significant challenges in real-world deployments where sensor data is often incomplete, noisy, and subject to nonstationary demand patterns. Traditional rule-based controllers lack flexibility under unpredictable conditions, while deep reinforcement learning (DRL) approaches, though powerful, require extensive training data and suffer from "black-box" interpretability issues, making them difficult to audit. Furthermore, real-world IoT sensing at intersections is compromised by sensor occlusion (e.g., by heavy vehicles), adverse weather attenuation, and stochastic infrastructure disruptions (e.g., accidents). The paper addresses the need for a traffic signal controller that remains effective under these noisy, partially observable conditions while offering a traceable decision-making process.

2. Methodology

The authors propose an Active Inference (AIF) controller for a four-arm signalized intersection. Unlike standard control paradigms that separate perception and action, AIF frames both as a unified process of minimizing Expected Free Energy (EFE).

Core Framework

The agent operates at discrete time steps $t$ , maintaining approximate posterior beliefs $q(s_t)$ over hidden states $s_t$ (traffic congestion levels) based on observations $o_t$ . The policy $\pi$ selects actions (phases: North-South or East-West green) to minimize EFE, defined as:
$G(\pi) \triangleq -[PV(\pi) + \lambda EV(\pi)]$
Where:

Pragmatic Value ($PV$): Rewards observations matching preferences (e.g., low congestion, low emissions).
Epistemic Value ($EV$): Rewards observations that reduce uncertainty about the hidden states (exploration).
$\lambda$ : A weighting constant balancing the two values.

System Instantiation

State Space: Hidden states are discretized into six congestion levels (from "very-low" to "jam") for each direction (NS and EW).
Observations: Vectors containing noisy vehicle counts, cumulative CO2 emissions, and bus counts. The observation model is a state-conditional multivariate Gaussian.
Transition Dynamics: Two manually specified transition matrices ( $T_{red}$ and $T_{green}$ ) encode the intuition that green phases reduce congestion while red phases increase it.
Noise Modeling: The experimental environment simulates realistic IoT failures, including:
- Sensor Occlusion: Heavy vehicles blocking detection.
- Weather Attenuation: Reduced detection probability during adverse weather.
- Stochastic Accidents: Random blockages freezing traffic flow.

Comparison Baselines

The AIF controller is benchmarked against:

Rule-Based Heuristic: A priority-scoring system using fixed bonuses for buses and main-road bias.
Deep Q-Network (DQN): A standard DRL agent trained with experience replay and $\epsilon$ -greedy exploration, optimized for the same reward function (minimizing idle time, emissions, and bus load).

3. Key Contributions

The paper makes three primary contributions:

Simulation Environment: Design of a SUMO-based simulation environment modeling noisy, nonstationary IoT sensing conditions, including sensor occlusion, weather attenuation, and stochastic accidents.
AIF Controller Development: Implementation of an AIF controller that dynamically selects signal phases by minimizing EFE over Gaussian beliefs regarding congestion, emissions, and public transport presence.
Comprehensive Evaluation: A rigorous comparison across four scenarios of increasing environmental disturbance (from clean data to full noise/accidents) and 100 independent random seeds, analyzing trade-offs between efficiency, emissions, and policy stability.

4. Experimental Results

Experiments were conducted over 3,600 seconds of simulation time per scenario. Key findings include:

Performance in High-Disturbance (Scenario 4):
- Idle Time: AIF achieved the lowest cumulative idle time (56,977 s) compared to DQN (71,741 s) and the rule-based baseline (72,247 s).
- Emissions: AIF produced the lowest total CO2 emissions (29.12 kg) versus DQN (30.56 kg) and the rule-based approach (31.53 kg).
- Robustness: AIF maintained a lower mean idle time than DQN even in the noisiest scenario, whereas DQN performance degraded significantly under full disturbance.
Performance in Low-Disturbance (Scenarios 1–3):
- DQN generally outperformed AIF in Scenarios 1 and 3 (e.g., 5,305 s vs. 7,287 s idle time in Scenario 3), likely due to scenario-specific pretraining.
- AIF outperformed DQN in Scenario 2 (7,341 s vs. 8,385 s).
- In all scenarios, both AIF and DQN significantly outperformed the rule-based baseline (roughly halving cumulative idle time).
Trade-offs:
- Bus Priority: AIF had a lower bus priority service rate (80.5–82.0%) compared to DQN (84.6–89.7%) and the rule-based system. This is attributed to the pragmatic value function prioritizing overall congestion reduction over specific vehicle types, as buses constitute a small fraction of total traffic.
- Phase Switching: AIF switched phases more frequently (261–278 switches) than DQN (177–266) or the rule-based system, indicating a more dynamic response to uncertainty.
Stability: AIF demonstrated lower variance (standard deviation) in idle time across seeds in Scenarios 1 and 2, while DQN showed tighter distributions in Scenarios 3 and 4.

5. Significance and Claims

The paper claims that Active Inference offers a principled alternative to deep reinforcement learning for adaptive traffic control, particularly in environments characterized by partial observability and nonstationarity.

Zero-Shot Generalization: Unlike DQN, which requires extensive pretraining for specific scenarios, the AIF controller operates without a pretraining phase. Its ability to match or exceed DQN performance in noisy environments without scenario-specific training suggests that the epistemic value term in the EFE objective provides a structural advantage for handling uncertainty.
Auditability: A significant advantage highlighted is the traceability of the AIF pipeline. Every stage—from posterior beliefs to the specific components of the EFE (pragmatic vs. epistemic)—is explicit and inspectable. This addresses the "black-box" verification challenges associated with DRL, allowing operators to understand why a specific phase was chosen.
Viability: The results suggest that AIF is a viable alternative to reinforcement learning for adaptive traffic control, capable of balancing multiple objectives (delay, emissions, public transport) while remaining robust to sensor degradation and stochastic disruptions.

The authors conclude that while AIF trades off some bus priority and increases phase switching frequency, its robustness in noisy, nonstationary IoT environments and its interpretability make it a compelling approach for next-generation intelligent traffic systems. Future work is proposed to extend the framework to multi-intersection networks and hardware-in-the-loop deployments.