How Human Problem-Solving Can Inspire Better AI Scheduling
SURF research considers how people plan their day to create more efficient calendar management tools
By Megan HarrisMedia Inquiries
- School of Computer Science
When scheduling an appointment, Carnegie Mellon University junior Elchanan Haas said most people make decisions quickly. How long will the activity take? Where will it occur? Is travel time needed? Is a specific day or time better for other attendees? They check their calendars, and slot tasks wherever they can.
That may or may not leave time for necessary extras, like dropping a letter in the post box or snagging coffee and groceries. People calculate whether they have time for these tasks innately, said Haas, who is studying computer science.
"But their quick choices don't always lead to the most efficient schedules," he said.
Maybe they forget a Thursday meeting is in the same neighborhood where they want to drop off dry cleaning. Or a new activity pops up that requires reordering a whole day.
"We want AI systems to handle these high-level scheduling tasks efficiently, and in ways that humans would appreciate and understand," Haas said.
Supervised by Stephanie Rosenthal, a CMU alumna and assistant teaching professor in the School of Computer Science, Haas spent the summer building on previous studies to create heuristics, or problem-solving strategies, that mimic and ultimately improve on human decision-making.
In a human research study last year, Rosenthal and computer science senior Poon Vichivanives — along with fellow CMU alumnae Laura Hiatt, now with the Naval Research Laboratory, and Elizabeth Carter, who serves as a project scientist with CMU's Robotics Institute — provided subjects a map and list of activities and asked how they would fit new tasks into the day.
The group found most participants scheduled tasks in either the first available block of time, ordering them consecutively from morning to night, or in the longest available free period.
By contrast, existing AI solutions consider every possible way to add the tasks into the schedule, Rosenthal said, which is time consuming when a schedule is already packed or the route is flexible.
"For example, there are lots of places to get a coffee. Lots of mail boxes near your workplace. Lots of places to acquire a pen, you know, if you need some new ones," she said. "When we look at the possibility of using AI systems to save human time, we have to think about how and when they would want to make these choices."
Vichivanives, who received his own SURF grant last year, blended human strategies and existing AI approaches to create a new set of algorithmic rules. This summer, Haas expanded that work to allow for calendar reorganization, which create much more efficient schedules but take time to compute. Haas also added heuristics that can schedule multi-part tasks, like picking up an item and delivering it to a new location at a later time.
"This could someday be adapted in a number of ways," Haas said. "Delivery companies are using algorithms to schedule their car fleets, but this complex mixture — pickups, deliveries, appointments, tasks — there's nothing on the robotics market today able to consider and efficiently schedule all of those activities at the same time."
The next step, Rosenthal said, is seeing how humans respond to that rescheduling.
"Just because a computer can find a better, more efficient itinerary for you, should it? Does shaving some time off really make my day better if I'm confused about where I'm going?" she said. "But we're not asking for those opinions just yet."
Vichivanives, who has remained involved through weekly check-in calls during a West Coast internship with Amazon, said he's pleased with the progress and hopes to rejoin the project team this fall.
Rosenthal, who earned her bachelor's, master's and doctorate all from CMU, said she's enjoyed watching the pair work.
"Research with professors was always my favorite part of being a student here. It's really exciting being on the other side of that now — to do the initial thinking on a very open problem, narrow down concrete questions or hypotheses to answer, give it to students and see what they can do."