Generalization and exploration in novel environments
17 Nov 2022 (Thu)
via Zoom
Mr Jaeho Choi, University of Pennsylvania

This study sheds light on the generalization problem in the context of strategic decision-making: how to generalize from prior experience in order to deal with new situations that are similar but different from the past? In particular, the study addresses a set of interrelated questions on learning and decision-making in novel environments: when presented with a stream of novel opportunities that generates uncertain outcomes, how do decision-makers infer the value of the opportunities using their prior experiences, and how do they decide to capture the opportunities based on their evaluations? Using a web-based experiment, which presents a task setting in which one either seizes a unique opportunity or not, the study provides novel findings about generalization and risk-taking. When holding positive expectations about novel opportunities, the decision-makers commit to the opportunities almost irrespective of the magnitude of the expectation. In contrast, when having negative expectations, they execute the reward opportunities to a noticeable degree, in proportion to the magnitude of their expectation. We interpret this pattern of behavior as a novel variant of the familiar exploration-exploitation tradeoff. In the positive domain of expectations, there is not an exploration-exploitation tradeoff − merely the incentive to “exploit.” In contrast, in the domain of negative expectations, there is a motivation to learn about the opportunity space as well, which leads the decision-makers to “explore” by taking risks.

KEYWORDS: Learning, Decision-Making, Generalization, Risk Taking, Experiment