We use cognitive models to apply the wisdom of the crowd to three sequential decision making problems: bandit problems, optimal stopping problems, and the Balloon Analogue Risk Task (BART). In each of these problems, people make a sequence of choices under uncertainty, with individual differences in decision making that depend on different attitudes toward risk. Each of the problems also has a known optimal decision-making strategy. Standard methods for the wisdom of the crowd, based on taking the modal behavior, are generally not applicable to these problems, because of their sequential nature. For example, the state-space of a bandit problem can be so large that, even for a large crowd of people, there will be game states that no individual encountered, and so there is no behavior to aggregate. We solve this problem using cognitive models, and inferring individual-level parameter values that predict what each individual would do for each sequential decision. The mode of these model-based predicted decisions then defines a crowd decision whose performance can be compared to individual and optimal performance.
A classical utilitarian perspective on public decision-making is that the public should choose an alternative that maximizes some function over the utilities of individuals, called a social welfare function. If the utility function for each decision-maker is cardinal and comparable with other decision-makers, then there are many valid social welfare functions. We examine the problem of finding a social welfare function that implicitly best fits society’s values, where we assume that a group's decision problem is the same as maximizing an unknown social welfare function implicitly held by the group. While there is a wealth of literature on possible social welfare functional forms based on different assumptions about the preferences of individuals, we propose a new empirical approach that approximates a group's social welfare function based on both individual preferences and group voting behavior. We test the approach's ability to promote compromise on climate-related energy policy among Pittsburgh residents who are plan to vote in the 2020 Democratic Primary. Our three-stage research design first elicits individual preferences, then uses a mean-variance algorithm to approximate a welfare function that fits group voting behavior, then finally makes a recommendation for the group based on that function. In addition to testing whether a group's social welfare function can be learned from individual choices, this study provides a roadmap for solving group recommendation problems more broadly.
Dr. Scott Brown
Dr. Ami Eidels
Dutch auctions are used in many industries. Goods are initially offered at a high price, which is gradually lowered until the first bidder accepts it. Bidders trade certainty and price: early bids secure the sale, but overpay; later bids are cheaper, but risk losing out to another bidder. We used group-based laboratory experiments to investigate decision-making in Dutch auctions. We developed a model for bidding in Dutch auctions, based on a dynamic extension of Prospect Theory. At each moment of the auction, the buyer is faced with a decision that can be framed as classical Prospect Theory: a certain option (buy now!) or a risky option (wait a little longer for the price to fall, and hope that no-one else buys before then). We show that this model reproduces the basic phenomena of the task, and also provides a useful framework for investigating interesting questions about auction psychology. We also discuss extensions to data from real Dutch auctions.