Possible ways to learn and represent human utility

Yao Yuanhang

Keywords: utility

Abstract: As described in the problem statement, the human utility function is internal to humans, lacks meaningful units of measurement, and varies across individuals. Estimating human utility for different tasks using a general computational framework remains challenging. In the following article, I will explore the characteristics of Preference-Based Reinforcement Learning, Expected Utility Theory, and conventional computational models such as decision trees and Bayesian networks, along with specific methods for learning and representing human utility, and their advantages and disadvantages in data collection, generalization, and efficiency.

Submission Number: 175