Snow Ploughing and AI: A Case Study in Decision-Making
In late January, the AiD – Center for AI in Decision Making at Norwegian University of Science and Technology(NTNU) in Trondheim was officially opened in the presence of the Minister for Research and Higher Education and the Minister of Digitalisation, Equinor Executive Vice President for Energy and technology transformation, NTNU pro-rector for research, and SINTEF CEO. The opening marked a significant national investment in trustworthy and robust artificial intelligence — AI that does not merely predict, but makes sound decisions under uncertainty.
At its core, AiD addresses one of the most challenging questions in modern AI:
How can machines make reliable decisions in complex, uncertain, real-world environments?
A vivid illustration of this challenge is the snow ploughing problem.
The Snow Ploughing Problem: A Black Square with White Squares
Imagine a black square — a simplified map of a city. On it lie white squares: roads covered by snow. The task is simple to state but difficult to solve: clear the roads efficiently.
Behind this image — reminiscent of “Vei i snø skog” (road in snowy forest) — lies a deeply complex decision problem:
- Weather forecasts are uncertain.
- Snowfall evolves dynamically.
- Some roads are critical (hospitals, emergency routes).
- Resources are limited.
- Decisions must be updated in real time.
This is not a static optimization problem. It is a sequential decision-making problem under uncertainty.
And this is precisely where probabilistic AI and its adaptation and integration into the decision processes become important.
Probabilistic AI: Modeling Uncertainty Explicitly
Real-world decision problems are never deterministic. Weather forecasts are uncertain. Traffic evolves unpredictably. Infrastructure systems are noisy and partially observed.
Probabilistic AI provides the mathematical tools to represent and reason about uncertainty explicitly. Instead of producing a single prediction, probabilistic models produce:
- Distributions over possible outcomes
- Quantified confidence levels
- Risk-aware decision criteria
This is crucial in snow ploughing: a strategy that is optimal for one weather scenario may fail catastrophically in another. Robust decisions must account for uncertainty, not ignore it.
At AiD, probabilistic modeling enables:
- Risk-sensitive optimization
- Bayesian decision frameworks
- Stochastic control methods
- Reliable forecasting integrated with decision layers
In short, probabilistic AI allows systems to know what they do not know — a key requirement for trustworthy AI.
Reinforcement Learning: Learning to Act
While probabilistic AI models uncertainty, reinforcement learning (RL) provides a framework for learning optimal actions through interaction.
Snow ploughing is inherently sequential:
- A plough clears one road.
- That action changes traffic patterns.
- Weather continues to evolve.
- New decisions must be made.
Reinforcement learning formalizes this as a Markov decision process, where an agent learns a policy that maximizes long-term reward.
RL is powerful because:
- It optimizes sequences of actions, not single decisions.
- It adapts to dynamic environments.
- It balances short-term costs with long-term benefits.
In infrastructure management, energy systems, logistics, and autonomous operations, decisions are rarely isolated. They are part of an evolving system. Reinforcement learning provides the tools to navigate such complexity.
However, AiD emphasizes that RL must be safe, robust, and grounded in structure. This is where integration with probabilistic modeling and physics-informed methods becomes essential.
Physics-Informed and Structure-Aware AI
A related research activity aligned with AiD’s vision is the PhysML project and upcoming workshop in Oslo, which focuses on physics-informed machine learning.
In many decision problems — including snow management, infrastructure systems, and energy networks — decisions must respect physical laws and engineering constraints.
Purely data-driven AI may produce impressive predictions but can fail when extrapolating beyond observed data. By incorporating:
- Physical constraints
- Geometric structure
- Mechanistic models
- Stability guarantees
AI systems become more reliable and interpretable.
This integration is central to AiD’s philosophy: decisions should be data-informed, uncertainty-aware, and physically consistent.
A National Commitment to Decision Intelligence
The January opening of AiD — attended by government representatives and research leaders — signals Norway’s recognition that AI must move beyond prediction toward responsible decision intelligence.
The center builds bridges between academia, public institutions, and industry, targeting applications such as:
- Infrastructure management
- Energy systems
- Autonomous operations
- Climate adaptation
- Public planning
From a black square with white squares to complex national systems, AiD represents a shift toward AI that acts responsibly in uncertain environments.
In a world shaped by climate variability, technological complexity, and societal risk, decision making under uncertainty is no longer optional. It is essential.
AiD aims to place Norway at the forefront of this transformation — developing AI systems that do not merely compute, but decide wisely.