Year: 2026
Type: dissertation
Abstract: We explore the creation of human models of poker players with a combination of deep neural networks and regularized search. We begin with behavioral cloning and explain our neural architecture choices. Next, we qualitatively show that our model can express a variety of human-like behaviors. Finally, we use the aforementioned model combined with regularized search to understand the tradeoff between human-likeness and policy strength in poker. In particular, we find that we can increase the human-likeness of a near-equilibrium policy for a low cost in exploitability, or increase the strength of a human-like policy for a low cost in human prediction accuracy. This thesis lays groundwork for using computers to create human-like opponents to train against.
Author: Chen, Nathan
Institution: Massachusetts Institute of Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7202060719
