Files in this item



application/pdfECE499-Fa2018-zhang.pdf (951kB)
(no description provided)PDF


Title:Realization of selfish learning
Author(s):Zhang, Wenxian
Contributor(s):Varshney, Lav R.
Degree:B.S. (bachelor's)
Subject(s):social learning
selfish learning
Markov decision process
reinforcement learning
Abstract:This thesis provides an overview and experiments/simulations to verify improved efficiency and accuracy of social learning through so-called selfish learning that revises traditional social learning patterns (Raman and Pattabiraman, 2018). We begin with introducing new concepts of social learning and the theoretical basis for utilizing selfish learning. Selfish learning is based on the concept that each agent will know decisions made by previous agents while receiving private signals. The agent can choose to declare and pass or declare and stop under the given information. There are two settings for selfish learning: 1) finite horizon, and 2) infinite horizon. For finite horizon, we develop crowdsourcing experiments with a finite number of workers to test whether agents will respond according to reward maximization. The task is basically a ball game with different situations of varying combinations of private signal and public signal; workers give their choices in different situations. Additionally, we determine an appropriate experiment sample size by Monte Carlo simulation. For infinite horizons, we use reinforcement learning to simulate a Markov decision process formulation. We find that by informing decisions made by previous agents and receiving private signals, the accuracy of answers from crowdsourcing will be improved and the cost will be decreased.
Issue Date:2018-12
Genre:Dissertation / Thesis
Date Available in IDEALS:2019-09-10

This item appears in the following Collection(s)

Item Statistics