AI-Powered Curriculum Optimization
Navigating Complexity in Educational Planning
The AI discussions in education are centered around tactical classroom approaches, but I think that’s scratching the surface. The more important long-term benefits of AI might be as a helper to optimize learning and administration.
Many years ago, I was involved in a series of projects in support of the U.S. Transportation Command, the portion of the military that plans all their movements. Logistics and transportation aren’t the glory roles in the military, but they address one of the most important and difficult tasks. Unlike commercial shipping, military operations face rapidly changing scenarios and evolving success metrics. They must balance short-term efficiency with long-term sustainability, considering factors like personnel burnout and equipment maintenance in a constantly shifting global landscape. The future uncertainty demands adaptable, resilient planning.
Schools face extremely difficult optimization problems too, and they are some of the most natural and important uses of AI in education.
How Optimization Can Help
Optimization is the process of finding the best, or at least the ‘good enough’. It is infused in a ton of life’s most important and difficult challenges, including at school. Even a relatively simple scheduling challenge, for example, can be difficult for human brains to fully grasp.
That’s where AI comes in. One of the big benefits of AI is it can chew on more information than we can. Except AI can’t solve difficult optimization problems on its own. It’s one of the problems that is shown to benefit from combinations of AI/algorithms and people.
In the late 2010s, Professor Dimitris Bertsimas and his students at MIT built an algorithm to optimize school bus routes for Boston Public Schools. Despite the promise of the algorithm, implementation faced several real-world obstacles. These included dealing with Boston's complex street network, accommodating various school start times, and managing parental expectations and concerns. The algorithm found more efficient bus patterns, but changing those routes in practice also required social evolution, insertion of additional values not considered in the math, and anticipation of second-order effects. The research showed Boston Schools could save millions of dollars by staggering school start times, but doing so affected more than the district’s efficiency, it affected the parent and employee lives too.
While the full optimization couldn't be implemented as originally envisioned, aspects of the research were used to improve Boston's school transportation system. In 2017, the school district reported saving about $5 million by using insights from the study to make route adjustments. The work was not a failure because the AI couldn’t do it all; it was a success because the combination of humans and AI could do what neither could on their own.
Getting closer to ‘best’ is a common challenge, but optimization problems can be very complicated, forcing out brains into shortcuts. We must either ignore some factors, which usually oversimplifies complex problems, or break the challenge into smaller chunks. It’s too much for our brains to manage. Somebody scheduling classes can’t simultaneously consider everyone’s schedule, preferences, policies, and personal constraints. A simple scheduling problem, like arranging just 10 classes into 10 time slots, already has over 3.6 million possible solutions. Usually the problem is broken into smaller pieces for manual analysis. Maybe the schedule is made by prioritizing seniority, then violations of policy or the need for human cloning are identified, and then the solutions are iterated until conflicts go away.
AI-assisted optimization might be a giant enabler of educational change. Modern higher education scheduling systems employ AI and other algorithms already, but that’s just a fraction of the complication in running an educational institution. Breaking the big picture challenge of optimizing the school into pieces via departments necessarily means foregoing opportunities that don’t fit neatly into the boxes.
Humans need to have it broken into simpler pieces because we can’t chew on the bigger issue without brain melt. Not so for AI, but optimizing complex systems is such a difficult problem that there are too many possibilities for the AI too. The AI-human synergy is powerful though, because people are good at quickly ruling out illogical or unlikely solutions, quickly constraining the possibilities to a small enough set for AI processing.
Optimizing an educational operation is a gnarly challenge that exceeds the difficulty of many logistics operations. Humans are reactive beings; packages are not. Success of the education system isn’t as easily measured as shipment time and cost. Yet, even with those difficulties, figuring out how to use AI on educational optimization problems may be the one of the biggest ways we can improve the system.
Curriculum Design as an Optimization Problem
Anyone who has developed even a moderately complex curriculum had to simplify the problem considerably. Even still, it’s a mentally taxing activity. Â
When the broadest view is taken, the scope of things that matter to curriculum is huge. Even inefficient school bus routes could have an impact on student learning, perhaps through additional student fatigue, that should ideally be accounted for in designing learning pathways.
Let me scale this back a bit, ignore school operational factors, and just focus on optimizing learning through course selection, design, and delivery. That’s not much less of a headache. A ton of factors matter, and they relate to one another in complex ways. Some goals are conflicting. And the system is dynamic; what worked before might not work now, at a minimum because the world outside the school, and the students and employees within it, are different than before.
Optimizing curriculum design is like trying to solve a Rubik's Cube where the colors keep changing. Just when you think you've aligned one side (say, content coverage), you realize it affects another (like student engagement or time constraints), and the educational landscape is constantly shifting beneath your feet.
Designers don’t try to tackle anything near the full complexity of the curriculum design problem. Instead, they break it into courses, units, and modules, each with short-term learning objectives, but with the consequence that long-term learning, such as for durable skills, are not optimized.
There may be other simplifications. Perhaps one pedagogy or assessment paradigm is favored to ease design workload and create consistency, but foregoing the benefits of varied pedagogy and assessment. Maybe curriculum solutions are designed to what teachers know how to do, but an incremental curriculum change plan could enhance skills without overwhelming educators. Or the curriculum is written generically for lots of schools, but each school and class would get more benefit from customized solutions.
Curriculum designers take what is one giant, inter-related hairball of a problem and make it simpler. In doing so, understanding the path for long-term learning is severely hampered. This affects the fostering of cross-cutting durable / 21st-century skills like critical thinking, interpersonal skill, and creativity most of all. It’s not the designer’s fault. Some of them are amazing thinkers. It’s just that they’re human.
Generative AI (GenAI) isn’t necessarily the type of AI that should be used for this problem, but it may be able to help in several ways. I expect the key is to talk to it in the language of optimization, not conventional curriculum and instructional design lingo. Admittedly, I don’t know how well this will go. I’m designing a K-12 AI Education Curriculum now, and GenAI keeps pushing me into a hierarchical design process that leads to module-specific optimizations that short-change broader objectives. I expect the value of genAI, in what will likely be a multi-AI suite of assistants, is in characterizing squishy qualitative factors in a mathy way that other AI can chew on.
To my knowledge, there aren’t AI products on the market to help curriculum design at this grander level. There are a few research papers that couch curriculum as optimization, but that framing hasn’t seemed to catch on.Â
There will be AI helpers for various tough optimization challenges, just like there are in logistics. The potential solutions are too ripe to ignore. The combination of AI and people could revolutionize how curriculum is made.
Your school can get ready for these helpers in a few ways:
Define the relevant information: Start playing around with AI modules that are increasing built into Learning Management Systems (LMS), either as primary options or as plug-ins. It will help you understand how your existing curriculum can be analyzed with GenAI, and what important contextual information might be elsewhere. Silo’d data systems are still an issue.
Judge ‘Better’: Every curriculum design paradigm I’m aware of starts with the educational goals and tries to ensure traceability of details back to those goals. However, the goals are understandably squishy, high-level, and likely tradeoff with one another. Optimization needs more specificity. Rather than endless debates that go nowhere about what school outcome is best, I suggest you converge on specific goals by rating hypothetical outcomes. Put comparisons of outcome measures in front of decision makers and ask which is better (we’re more sensitive to relative ranking). Do that enough times with various outcome combinations, and a statical profile of best will begin to emerge.
Use GenAI for ideation on human/AI task decomposition: The art in human-AI collaborative optimization is in figuring out how to break up the problem so that the talents of both are applied. That can mean restructuring existing processes. GenAI isn’t going to give you a good plan on its own, but it can help. I like asking it about how analogous challenges have been approached in other fields, which can stimulate thinking.
Though on the horizon, AI will help optimize the grander goals of education. In the meantime, the optimization mindset shift can begin.
AI is a paradigm changer, and it takes a while to grasp that what held people back from taking other solution paths may no longer be as big a constraint. We do what we do the way others have done it without considering another way because the prior constraints were baked in, barely conscious. Then we’re surprised when the new paradigm emerges.
This paradigm change is predictable. AI will be able to help curriculum design, not just lesson planning. That will require as much human creativity as computing.