Complex data-rich endeavors like predicting climate change or developing a new heavy-duty material involves running many simulations and comparing their results with observations. Each simulation or experiment is an “information source,” whose use has its own pros and cons.
“In trying to optimize something using many sources of information, each source has a different cost to run, both in time and resources, and a different accuracy,” says SFI External Professor David Wolpert, an algorithm designer at Los Alamos National Laboratory.
Traditionally, scientists use their intuitions to choose from information sources on the fly. Wolpert wants to let machines do it instead. “Humans aren't designed for this,” he says. “Statistical techniques may be far more powerful.”
A dozen researchers in representing MIT, Stanford, Sandia National Labs, and the Air Force gathered at SFI recently to consider new directions for Multiple Information Source Optimization (MISO). As part of the working group, they explored means of finding statistical relationships between cheap but inaccurate sources and accurate but expensive sources.
“The right knowledge allows you to use cheaper, less accurate sources knowing how they correlate to the results of better ones,” says Wolpert, who held a similar working group in July 2012, during which organizers introduced standard MISO procedures – what had so far been used in aerospace engineering – to other research communities.
At the follow up meeting in November they introduced semi-supervised machine learning: an approach to optimize the optimization process by explicitly comparing an information source’s cost to the degree it improves the design.
The advances are poised to make a tremendous difference in fields requiring massive simulations, say Wolpert, such as determining the human dosage of pharmaceuticals, fuel yield of algal farms, magnitude of climate change, and severity of extreme weather.
Participants recently presented their findings at the Society for Industrial and Applied Mathematics conference in Boston.
More about the SFI working group