Australasian Mathematical Psychology Conference 2019

The value of predictive information in decision-making under uncertainty

Ariel Goh
School of Psychological Sciences, Monash Institute of Cognitive and Clinical Neurosciences
Daniel Bennett
Princeton University
Stefan Bode
The University of Melbourne
Trevor T-J Chong
Monash University

Humans exhibit a drive towards acquiring information. Notably, studies of humans and non-human animals suggest that information is processed by similar neural circuits that underlie reward valuation. This project investigated how humans value information that predicts, but does not change, the outcome of an upcoming event (non-instrumental information). We conducted two experiments to examine the physical effort costs individuals are willing to incur for such information. Effort was operationalised as amounts of force applied to a hand-held force-sensitive dynamometer. In the first experiment, the amount of information available was held constant, and participants chose between exerting higher effort levels to obtain predictive information about a lottery outcome, versus exerting minimum effort and foregoing such information. Results showed that participants willingly exerted effort to obtain the information, but this effect declined as effort costs increased. In Experiment 2, we manipulated the amount of information provided at the start of each trial, and thus the amount of uncertainty participants experienced. Results showed that participants invested more effort for information when prior uncertainty was high (i.e., when the outcome was ambiguous) compared to when it was low (i.e., when the outcome was predictable). Bayesian model comparison using the Watanabe-Akaike Information Criterion revealed that subjective valuation of information was best modelled as a function of both effort costs and the magnitude of available information, where information was quantified as the degree to which information reduced residual uncertainty about the outcome. Model comparison also suggested that participants’ uncertainty was best modelled by the Rényi entropy of beliefs (a generalisation of Shannon entropy). Overall, these results suggest that information’s intrinsic value is based on its capacity to reduce uncertainty, and that this valuation is reflected in a willingness to trade off effort for information. This work helps explain the bias humans exhibit toward information acquisition, even when this is sub-optimal or inefficient.