When NASA scientists equip the Orion spacecraft with medical supplies for its manned mission to Mars, they will choose what to send in light of their best predictions of what the environment will be like there. Yet some of the most difficult scenarios for scientists to predict are those in which agents enter into unknown territories. Can scientists develop methods that will help NASA make better choices for equipping its mission? A February SFI working group, “Lookahead optimization in artificial and natural systems,” brings together scientists from diverse fields to develop better quantitative models of optimal decision making. The interdisciplinary working group was conceived by SFI and MIT Postdoctoral Fellow Brendan Tracey, SFI Professor David Wolpert, and SFI Professor Mirta Galesic, who is Cowan Chair in Human Social Dynamics.
Tracey and Wolpert began planning the group when they discussed the limitations of current optimization algorithms, which tend to focus on immediate payoffs, rather than on the relative benefits of learning new information. “The value of lookahead optimization,” according to Wolpert, “is that it gives us a way to formalize how agents gather and then exploit information. When aerospace engineers test airplane wings, they should not choose what to test next without accounting for what they will learn in their initial test, as the information they learn will affect subsequent choices of what to test.” Lookahead optimization allows scientists to account for this kind of learning.
Galesic became involved in the workshop when she recognized that lookahead optimization may help us understand some seemingly odd patterns in individual and social decision making. “It can be hard for us to see how a decision might actually be optimal for a set of actors. Sometimes what does not look optimal — say, delaying an important decision rather than choosing what seems like a good solution right now — might actually make sense in a lookahead framework, which accounts for the long-term consequences of immediate choices.”
While Galesic hopes that the workshop will help her see where lookahead optimization might be used to understand and predict human decisions, Tracey and Wolpert hope to learn more from Galesic about how heuristics that humans and other animals use relate to engineering design. For Tracey, the working group is an occasion to clarify “the meeting ground between mathematical decision models and patterns in biological, social, and artificial systems.” The workshop takes place at SFI from February 21-22, 2018.