For nearly two decades, neuroscientist Karl Friston has argued for a strikingly compact theory of what brains are for: not passively receiving the world, but constantly generating predictions about it and working to minimize the gap between what they expect and what they sense. He calls the mathematical quantity being minimized “free energy,” borrowed from statistical mechanics but repurposed here as a measure of surprise. Perception, learning, and even action all fall out of the same single imperative, in Friston’s account — an organism resists disorder by continually updating its internal model of the world, or by acting on the world to make its predictions come true. In May 2026, a team of engineers took that theory and built a traffic light controller out of it. The interesting part isn’t that it worked. It’s what that success actually reveals about the theory itself.
Scientific Foundation
The free energy principle, formalized by Friston starting in 2006, casts perception as a hierarchy of prediction and error-correction: higher levels of a cortical hierarchy generate predictions about lower-level sensory input, and mismatches — prediction errors — flow back upward to update the model. Extended into “active inference,” the same framework accounts for action too: an agent can reduce surprise either by revising its beliefs to match reality, or by acting to change reality so it matches its beliefs. The mathematics underlying this, variational free energy minimization, is drawn from Bayesian inference and is genuinely unifying in scope — the same objective function has been used to model everything from single-neuron dynamics to full-body motor control to, as of this year, decision-making in repeated social games.
That last extension is what made the traffic light experiment possible. A team led by Dénes Tóth built an active inference controller for a four-way signalized intersection, one that dynamically selects which direction gets a green light by minimizing expected free energy over its beliefs about congestion in each direction, arriving at a decision process the researchers describe as fully traceable — every choice can be audited back to the beliefs and predictions that produced it. Tested in the SUMO traffic simulator against both a simple rule-based heuristic and a deep Q-network, a standard reinforcement-learning approach already used in real adaptive traffic systems, the active inference controller matched or outperformed the deep-learning baseline specifically in the noisiest, most degraded conditions: occluded sensors, bad weather, unpredictable accidents. It did so with no pretraining at all, while the deep Q-network needed 500 to 1,000 simulated episodes of scenario-specific training before it worked well.
Cross-Domain Connection
It’s tempting to read this as proof that a theory of biological cognition has quietly infiltrated traffic engineering, revealing some deep kinship between brains and cities that nobody previously suspected. The more precise and, honestly, more interesting story is narrower than that. Active inference and reinforcement learning aren’t strangers meeting for the first time at a traffic intersection — they’re close mathematical relatives, and the relationship has been worked out in detail. Under maximum-entropy reinforcement learning, a well-established variant that rewards policies for maintaining exploratory randomness alongside pursuing reward, researchers have shown active inference and RL become formally equivalent, producing identical policies under appropriately matched priors. Broader reviews go further, showing that expected free energy minimization, expressed in variational form, subsumes classical reinforcement learning, optimal control, and Bayesian decision theory as special cases rather than standing apart from them. What active inference adds isn’t a foreign framework bolted onto traffic control — it’s a generalization of the reinforcement-learning math already used there, one that natively bakes in uncertainty and an “epistemic,” curiosity-driven term that plain reward-maximizing RL has to add on as an afterthought.
What Remains Undemonstrated
That reframing matters for how much weight this result can bear. The free energy principle’s status as a genuine, unifying theory of brain function remains actively contested within neuroscience and philosophy of science; critics have argued the framework is broad enough to resist falsification, even as supporters point to real neural evidence for predictive coding and precision-weighted attention. Treating a working traffic controller as a validation of Friston’s grand claims about biological self-organization would overstate what a simulated four-way intersection can actually demonstrate. It’s also worth being clear-eyed about the traffic result on its own terms: this is a single simulated intersection tested in software, not a deployed system managing real cars. Existing real-world adaptive traffic systems, built on older, classical control-theoretic logic, and various deep-RL pilots already installed in some cities, remain the operational standard; this active inference controller hasn’t yet been tested against real sensors, real weather, or real drivers.
Why It Matters
What the result does demonstrate, cleanly, is something more modest and more useful: that a framework built to explain cortical prediction-error minimization turns out to be mathematically compatible enough with reinforcement learning that it can be dropped into an RL-shaped engineering problem and hold its own, particularly where interpretability and robustness to sensor noise matter more than squeezing out the last percentage point of optimality. That’s a genuine, meaningful cross-pollination between neuroscience theory and infrastructure engineering. It’s just a more precise kind of meaningful than “your brain and a traffic light run on the same code” — it’s closer to discovering that two different fields, working on superficially unrelated problems, independently backed into neighboring corners of the same underlying mathematical space.
Human Dimension
There’s something fitting in that, even scaled down from the more dramatic version of the claim. A researcher trying to explain how a cortical column decides what it’s looking at, and an engineer trying to decide which lane gets the green light next, turn out to be doing recognizably related arithmetic — not because traffic is secretly conscious, but because “how do you act well under uncertainty with an imperfect model of the world” is, it turns out, a narrower mathematical question than either field would have guessed on its own.
Sources:
1. Springer Nature — “Free-Energy Principle and Predictive Coding: A Computational Theory Explaining Various Brain Functions” — https://link.springer.com/chapter/10.1007/978-981-95-1327-7_14
2. National Science Review (PMC) — “Bayesian brain computing and the free-energy principle: an interview with Karl Friston” — https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11060478/
3. arXiv — “Applications of the Free Energy Principle to Machine Learning and Neuroscience” — https://arxiv.org/pdf/2107.00140
4. arXiv — “Active Inference for Adaptive Traffic Signal Control in Noisy Nonstationary IoT Environments” (Tóth et al., 2026) — https://arxiv.org/abs/2606.13698
5. MDPI, Electronics — “Artificial Intelligence-Based Adaptive Traffic Signal Control System: A Comprehensive Review” — https://www.mdpi.com/2079-9292/13/19/3875
6. Ideasthesia — “Active Inference Agents vs. Reinforcement Learning: A Comparison” — https://www.ideasthesia.org/active-inference-agents-vs-reinforcement-learning-a-comparison/
7. PMC (National Institutes of Health) — “Decision, Inference, and Information: Formal Equivalences Under Active Inference” — https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12840411/
8. arXiv — “Active Inference in Robotics and Artificial Agents: Survey and Challenges” — https://arxiv.org/pdf/2112.01871
Idea originated at artificialideas.org. Article researched and written by Claude Sonnet 5. Published at artificialideas.org.
Artificial Ideas 