ePoster, COSYNE 2023
Information correlations reduce the accuracy of pioneering normative decision makers
Zachary Kilpatrick, Megan Stickler, Bhargav Karamched, William Ott, Krešimir Josić and Kresimir Josi
University of Colorado Boulder; Applied Mathematics; Carthage College; Florida State University
- Montreal, Canada
- Computational Neuroscience, Cognition and
Abstract
Decision making is essential to cognition, often requiring the integration of sequences of observations over time. Choice rules trigger decisions once sufficient evidence is acquired for a choice. Models often describe single accumulation-to-bound processes, but organisms often decide in groups or accumulate multiple information streams. A better understanding of naturalistic decision making thus requires probing the neural computations and behaviors of evidence accumulation processes involving the integration of multiple information streams. Normative models of information-sharing in multi-agent systems reveal recursive processes that prevent overcounting knowledge received from neighbors, building on log likelihood ratio (LLR) update models arising from sequential analysis [3,4]. Here, we consider a complementary question: How do correlations in the information collected by individual members of a group shape the relative accuracy of their decisions?
Our main finding is that individuals that make earlier decisions than other group members are less accurate when the evidence the members gather is correlated, even when all agents have identical decision criteria requiring the same total evidence. As a result, an agent's level of accuracy is ordered inversely by their decision time order (i.e., the nth decider is the nth least accurate). Joint dependence of each decider's accuracy on their own LLR and decision order arises due to correlations between the agents' evidence streams. This effect does not occur when agents make independent measurements of their environment, and is thus not due simply to early deciders having accumulated less evidence. The common evidence increment shared by all agents results in early decisions that are more likely to be based on faulty evidence. This finding reveals the importance of considering how information correlations common in nature impact decision accuracy. Our results give a normative account of reduced accuracy of hasty deciders even when all group members have identical decision criteria.
Topics
- Decision making, search World Wide
- evidence accumulation, search World Wide
- sequential analysis, search World Wide
- log likelihood ratio, search World Wide
- multi-agent systems, search World Wide
- decision accuracy, search World Wide
- decision time, search World Wide
- Normative models, search World Wide
- information-sharing, search World Wide
Details
- Session
- Poster Session I