|
Abstract
|
Using a computational approach, we investigate the optimal representation of subjective value in the resource-constrained brain in environments characterized by different reward distributions. We compare the optimal utility functions that emerge under two distinct objectives: first assuming that the objective of the chooser is the maximization of expected rewards and second assuming that the objective of the chooser is the minimization of the number of decision errors. Our results show that the optimal reward-maximizing encoding function depends critically not only on the underlying reward distribution but also on the level of noise in the nervous system’s encoding process. As expected, in a low-noise decisional system linear encoding optimally maximizes expected returns and minimizes errors. As noise increases, more complex encoding functions become optimal, such as sigmoidal or logarithmic, depending on the distributional structure of the environment. Finally, we quantify the expected payoff gains and error rate reductions that would result from relaxing the resource constraint in the brain i.e. from reducing the noise in value encoding. Our research provides a framework for understanding the link between reward encoding, environmental conditions, neural efficiency, and utility with implications for choice theory.
|