Using Small, Invisible Applications of AI to Power Large Changes
by Alice Lubeck
Content
Story by Alice Lubeck, UMass Amherst Public Writing Fellow
As a kid, Adam Lechowicz (he/they), a graduate student at UMass, dismantled their toys to investigate the invisible mechanisms that made them work. They carried this fascination of the little things that power larger systems with them to UMass where, as an undergraduate student, they majored in Computer Science and Political History. Their undergraduate studies showed them how technology has affected people throughout history, and this historical perspective motivated them to find ways to leverage modern technology to make society better.
Their passion for using technology to improve the human condition led them to a PhD program at the Manning College of Information and Computer Sciences where they design decision making systems with Artificial Intelligence (AI) to solve new problems. AI is a set of technologies that allow computers to reason, solve problems, and perform advanced tasks without human intervention. Lechowicz incorporates AI into complex decision-making systems to automate decision making and take the burden off of individuals while achieving multiple goals simultaneously. Specifically, Lechowicz focuses on real-time decisions, meaning decisions that need input every 15 minutes or less because of changing conditions.
The result of Lechowicz’s work is an invisible system that merges human and AI capabilities. Because their work is theoretical, they are focused on designing a system that can be applied to different problems, and Lechowicz is passion about applying this decision-making system in ways that can reduce carbon emissions.
“The sort of the overarching thing is really making decisions in real time, in an uncertain setting and [I’m] particularly motivated by the complexity and the challenges that climate change brings and the changes to the energy system.”
To describe these decision-making systems, Lechowicz gives an example of a system used to determine when to charge a battery. The first goal of the system would be to not overcharge and cause damage to the battery. The second goal would be to charge the battery when the electric grid is primarily powered by renewable energy to minimize carbon emissions associated with the task. Historically, systems designed to determine when to charge the battery would be conservative, meaning they would minimize the risk of overcharging the battery at the expense of maximizing clean energy use. While it is important to minimize a negative outcome, these historic systems would not be the most efficient.
To maximize performance of these decision-making systems, Lechowicz writes a set of possible algorithms, instructions designed to accomplish a task that optimizes decision making and then using AI to pick the best one. To do this, AI assesses outside data, like historical weather data, and uses that background information to predict which algorithm will be most efficient while avoiding negative outcomes.
There are three reasons Lechowicz’s work is particularly impactful. First, by combining human designed algorithms with a little bit of prediction by AI, the conclusions of the decision-making systems are understandable. Whereas an unrestricted use of AI could cause unintended consequences in an automated task because we do not know how AI makes decisions. Second, these changes do not need new infrastructure or heavy investment to be implemented. Meaning these decision-making systems are little wins towards energy efficiency, can easily be implemented and can accumulate into meaningful change. Finally, Lechowicz’s specifically addresses problems would be tedious or impossible for a human actor to engage with. This means energy efficiency can be achieved without too much human input.
Lechowicz, along with their advisor UMass Professor Mohammad Hajiesmaili, recently published in USENIX their work determining an algorithm to maximize efficiency and reduce carbon emissions when assigning jobs to data centers. As the United States' electric grid continues to incorporate renewable energy, AI can be used to solve problems that arise from that transition or other major infrastructure changes such as powering data centers. “There's going to be more of these problems, like the data center example, like the battery example that I've been giving, where a little bit of prediction...being able to learn a better algorithm that can do these things more effectively is going to, in aggregate, save a lot of, save people a lot of time and save, hopefully, the planet a lot of unnecessary emissions."