Professor of Computer Science
A Case for Reasoning
There’s no substitute for personal experience, at least according to conventional wisdom. But conventional wisdom just may have to change.
Problem solving has long been about remembering what you did when you ran across an issue in the past and either repeating the solution or modifying what was done before. Another option was to talk to other people to find out how they dealt with the issue and follow their advice.
But for David Leake, Executive Associate Dean and Professor of Computer Science at IU’s School of Informatics and Computing, and a number of other researchers worldwide like him, the future may hold better and more varied solutions thanks to artificial intelligence systems contributing the fruits of their experiences, by case-based reasoning.
“The idea is if you have a system that’s interacting with the world over an extended period of time, it’s going to capture a lot of experiences,” Leake says. “It can learn by storing the new cases and their solutions, to remember them later. Another other level of learning, on top of that, is learning to make better use of the cases it has. That might involve things like reorganizing the cases or learning new ways to adjust old solutions, to solve a broader class of problems in the future or avoid repeating mistakes. In order to do that, the system needs to be able to detect problems, and also to be able to explain what’s going wrong in order to fix it.”
Finding a way of efficiently sorting through all the cases that might exist in a system and distilling what solutions can be drawn from those examples has been a focus for Leake, and he has even taken the passage of time into account. In other words, what may have been true in the past for a solution may no longer apply.
“At what point does a past case become obsolete?” Leake asks. “Let’s say you’re a realtor, and you’re trying to assign a price to houses. You may remember a house that is a perfect match to the one you’re looking at now, but it was last on the market 20 years ago. You don’t want to set your new prices based on that. The question is both how a case-based system can distill the examples from all those years down to a manageable set, and how it can get relevant lessons from old cases. When should past cases be revised, not based on the similarity of the two instances – like “this house looks like this other house” – but taking temporal factors into account?”
Leake envisions future case-based reasoning systems interacting flexibly with people, to create better solutions as a team.
“Some of my work has been on an approach called conversational case-based reasoning, in which the system is guided by questions it asks the user” Leake says. “There’s a lot of reasoning that goes on under the hood about which questions to ask, to get to a solution as efficiently as possible. Then the system proposes suggestions and continues the interaction. The approach has been widely used in help desks. Some time ago we did a project on this for telemaintenance, with collaborators at (Crane Naval Base). Those systems have been very successful, and we can imagine building on that for much richer interactions for other aspects of the process, such as revising solutions.”
In fact, case-based reasoning systems can even anticipate problems before they arise.
“If a CBR system has a big library of cases,” Leake says, “it can use them not only to suggest, ‘Here is what you should be doing,’ but to help assess proposed solutions. If someone has a tentative solution, the system can look back and say, ‘I see you’ve proposed this solution, but it looks an awful lot like this solution, which ran into the following problems.’ The system’s experiences become a vehicle for anticipating problems and avoiding them.”
Leake, who has played a large role in establishing case-based reasoning as a major area of Artificial Intelligence, is excited by the explosion of data and how it can impact case-based reasoning in the future.
“In the very early days, people were doing small prototypes to illustrate particular research issues with very few cases,” Leake says. “I have a student now who is working with 100 million cases and looking at how to scale the methods up. There are staggering numbers of experiences on the Web. It’s a huge opportunity for systems that can harness what’s already there and keep on learning.”