Who, When, How (Not) to Imitate? The Role of Imitation in Collective Intelligence, and Its Implications on the Design of Socio-Technical Systems
Name
Choi-Eunseo-choie-sm-tpp-eecs-2024.pdf
Description
Thesis PDF
Size
9.48 MB
Format
Adobe PDF
Checksum (MD5)
3fd6aa3984f831920f208d7f8a1ea52e
Author(s)
Choi, Eunseo Dana
Advisor(s)
Hadfield-Menell, Dylan
Date Issued
February 2024
Publisher
Massachusetts Institute of Technology
Abstract
Humans collectively demonstrate coordination and progress on a massive scale, building, adapting, and thriving under the rules of different institutions. Researchers posit social learning as a mechanism for overcoming individual limitations, quickly adapting to environments, passing knowledge across generations, and enabling rapid cumulative cultural evolution. This thesis demonstrates how multi-agent learning (MAL) can facilitate counterfactual experiments that shed light on the performance of different social learning. Simulations present that the details of who, when, and how to imitate affect group fitness in distinct ways based on the size and homogeneity of the group: 1. unbiased imitation works well in homogeneous groups as long as there is a minimum age for agents to be imitated; 2. imitation strategies based on models’ complete action history instead of their recent actions, although similar, can attain very different levels of group fitness; 3. very high levels of imitation probability (up to 98% in some cases) may be efficient for group learning. Results from this thesis complement and contradict accepted results from the literature. By explicitly comparing the mechanisms that govern the success or failure of group learning, findings from multi-agent learning can provide essential guidance for the design of socio-technical systems.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Massachusetts Institute of Technology. Institute for Data, Systems, and Society
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
