These are the questions at the heart of the Research Semester Programme (RSP) Tailored Optimization, which started earlier this year. During its first meetings, researchers explored the obstacles they encounter when adapting algorithms to real-world problems. In workshops, they also examined ways to systematize the time-consuming and costly process of fine-tuning algorithms, or even to incorporate it into an automated workflow. At the final meeting on 21 and 22 May, the wrap-up meeting, participants will work on turning these insights into practical guidelines.
Challenge
“We work on basic, ready-made algorithms that can be used for all kinds of optimization problems,” says co-organizer Vanessa Volz of CWI’s Evolutionary Intelligence (EI) group. “In short, such an algorithm contains the steps needed to solve the problem.”
In theory, such an algorithm can be applied to many different situations. But, Volz explains, the real challenge is tailoring it in such a way that it produces the desired solution for a user’s specific problem. That requires an understanding of the problem context: what exactly does the user want to achieve with the algorithm? Which aspects matter most to them, and which are less important?
Asking the right questions
Volz gives an example: “Suppose someone sells eggs in five different sizes, each with its own price. They want to move to a simpler system with just two prices: price X for eggs measuring 50 to 53 centimetres, and price Y for eggs larger than 53 centimetres. Anything smaller than 50 centimetres is thrown away. I can design an algorithm that calculates the perfect price. But because I haven’t asked the right questions, I only discover after running the calculations that they actually still consider eggs of 49 centimetres acceptable and do not throw them away. Then I have to start all over again.”
“Because I need to uncover assumptions like these that a client does not state explicitly, a great deal of discussion is needed. That is how I can build a good representation of the problem in the computer. The challenge is knowing where to start when asking questions. For many researchers, that communication turns out to be a bottleneck.”
Role play
At the wrap-up meeting, the aim is therefore to bring together practical insights on tailoring optimization algorithms into accessible guidelines for researchers. The communication challenge – the social barrier between academia and industry – will also be addressed through role play.
Volz: “We are going to practise conversations. Someone plays a client who does not quite know what they want. How do you start that conversation? For instance, by offering two solutions and asking the client which one they prefer. In that way, we want to share the insights we ourselves have gained through trial and error, so that others do not have to go through that same long learning process.”
The next step, Volz continues, is to develop a more systematic solution, for example in the form of a book of tips and tricks. “If you’re stuck, build this – something along those lines. But for now, we are mainly collecting anecdotes to learn how to talk to the user.”
Photo: Shutterstock
