Paul Glimcher

Paul Glimcher

New York University

Optimal Utility

NYU Meyer Hall, 4 Washington Pl, room 636

Abstract

Using a computational approach, we return to Pascal’s question: What should the human utility function look like? Here we assume that humans have finite cognitive capacity, want to maximize an objective goal, but unlike earlier scholars we assume that humans can select the specific utility function they employ in any given environment. These assumptions allow us to characterize an optimal utility function. Our results show that almost every experimentally observed utility function can be described as a near optimal tool for achieving a maximization of objective gains (a goal abandoned since Bernoulli). In a high cognitive capacity decisional system, linear utility functions optimally maximize expected returns and minimize errors. As cognitive capacity decreases (internal noise increases), more complex encoding functions become optimal, but in a manner dependent on the structure of the environment. We also quantify the specific gains that result from adding more cognitive capacity and find that these gains are quite small as long as the utility function is correctly adjusted to compensate for losses in cognitive capacity. Our research provides a novel framework offers a foundational explanation for why human and animal decision-makers evolved the puzzling range of utility functions and risk attitudes that have been widely observed.