UNB News
News and stories from one of Canada’s top universities

UNB researchers have created a head-turning AI model that balances curiosity with efficiency, allowing it to refine its strategies in real time. Their recently published paper details how this system outperformed traditional models in managing pandemic policies by reducing health and economic risks. This patent-pending technology is already attracting interest in potential healthcare, energy and robotics applications.

A groundbreaking artificial intelligence (AI) algorithm developed at the University of New Brunswick (UNB) could transform decision-making in unpredictable situations, from healthcare to energy management.

The new approach, known as curiosity-based control, allows AI to seek out information while performing its tasks, an ability that sets it apart from traditional AI models.

“Most AI systems either passively learn from data or adapt based on what happens to them,” explained Dr. Jon Sensinger, an electrical and computer engineering professor and director of UNB’s Institute for Biomedical Engineering.

“Our algorithm does something different: It actively explores its environment to improve future decisions.”

The research, co-led by Dr. Sensinger and Dr. Juan Antonio Carretero, professor and associate dean academic in the faculty of engineering, centres on work by recent doctoral graduate Dr. Andrew Mathis (BScEng’16, D-TME’16, MScEng’19, PhD’24), who designed an AI system that balances this curiosity with efficiency.

The researchers recently published a paper entitled Beyond Adaptive Control: A Control Method for Nonlinear Systems With Uncertainties, Applied to COVID-19.

The study presents an advanced AI-based control system that thinks ahead instead of reacting. When applied to COVID-19 policies, it led to smarter, more innovative and effective decisions. The approach has potential applications in many fields where uncertainty is a challenge.

How did it start?

Inspired by his experience managing his passive solar greenhouse, Dr. Mathis initially conceived an AI greenhouse control system.

“There is a temperature control, a humidity control, lights, ventilation and a passive solar and geothermal system built into it,” Dr. Mathis said.

“I proposed researching that, but then we realized we could develop something more widely applicable than just greenhouses.”

AI that proactively learns

Unlike conventional AI models, this new algorithm proactively tests its assumptions. It carefully probes uncertain aspects of a system while still working toward its primary goal, whether saving lives, reducing costs or optimizing efficiency.

One great advantage is this AI runs efficiently with minimal computing power: Dr. Mathis ran it on his decade-old laptop.

The key innovation lies in how the AI balances two types of information: states, which are tracked and adjusted, and parameters, which are slower-changing variables in the system. The challenge was to make the AI curious about these parameters while still focusing on its primary task.

“What is cool is this control algorithm can prioritize both things: learning while saving lives,” Dr. Mathis said. “It could test things when stakes were low, and then it had more information to make better decisions when the stakes were high.”

A new approach

The research paper explores a new approach to managing systems where outcomes are unpredictable due to uncertainties.

It introduces an advanced control method called dual iterative linear quadratic Gaussian (iLQG) control, which improves decision-making compared to traditional adaptive control.

Adaptive control learns from past data but does not plan future actions to improve learning. Dual control, on the other hand, experiments to gather helpful information, improving long-term outcomes.

Dr. Sensinger compared this method to driving on an icy road.

“You could just drive normally, applying past knowledge about winter driving. Or you could wait until you skid at an intersection and then react,” he said.

“But a better approach would be tapping your brakes early, testing the road conditions in a way that helps you make better decisions as you go. That’s what this AI does.”

The study applies dual iLQG to managing COVID-19 restrictions, using data compiled during the pandemic. This approach helped the AI find the best mix of COVID-19 policies (such as lockdowns or mask mandates) by balancing public health and economic impact, both of which had been government concerns.

It improves existing methods by simultaneously handling multiple uncertainties, circumventing a challenge known as Bellman’s curse of dimensionality.

The dual iLQG controller outperformed adaptive control by 6.4 per cent in reducing COVID-19 deaths and economic costs, while adjusting policies based on new data over time—proactively reducing uncertainties rather than just reacting.

Unlike standard models, it also accounted for the impact of overwhelmed healthcare systems. It proved effective even when handling 16 uncertain parameters in a complex scenario.

Real-world applications

This AI could help governments and businesses make better decisions in uncertain situations, from pandemic responses to energy management and robotics.

The model has potential applications in diverse fields, such as:

• Smart grid optimization, improving energy distribution efficiency;
• Building climate control, making heating and cooling systems more adaptive;
• Healthcare rehabilitation, personalizing treatments for spinal cord injury patients; and
• Financial systems, helping businesses make more informed economic decisions.

The UNB team is already exploring collaborations, including a research project with the Stan Cassidy Centre for Rehabilitation and Praxis Spinal Cord Institute.

Early discussions with healthcare authorities and local companies suggest this AI model could be used to improve resource allocation in hospitals and government policy planning.

While it performed well in simulations, real-world applications still need fine-tuning to account for human behaviour and external factors.

Commercial potential and future research

With a patent pending on the technology, UNB’s Fulcrum initiative seeks industry partners.

Dr. Sensinger believes this kind of collaboration is key to unlocking its full potential.

“This is an incredibly powerful tool, but for it to have an impact, we need to work with experts in different fields: policymakers, healthcare professionals and energy analysts,” he said.

“The algorithm can suggest innovative solutions, but its strength lies in helping human decision-makers make better-informed choices.”

The researchers plan to refine the algorithm further, expanding its use in more complex, high-stakes environments. Dr. Sensinger sees possibilities for curiosity-driven AI, not just in automation but in scientific discovery itself.

“Curiosity is at the heart of all great research,” he said. “If we can build AI that embodies that same principle, we open the door to smarter, more adaptive decision-making in every field.”

Photo: Dr. Andrew Mathis (BScEng’16, D-TME’16, MScEng’19, PhD’24), who designed an AI system that balances curiosity with efficiency and needs so little computing power that he runs it on his laptop.