Can natural selection encode Bayesian priors?

Abstract

The evolutionary success of many organisms depends on their ability to make decisions based on estimates of the state of their environment (e.g., predation risk) from uncertain information. These decision problems have optimal solutions and individuals in nature are expected to evolve the behavioural mechanisms to make decisions as if using the optimal solutions. Bayesian inference is the optimal method to produce estimates from uncertain data, thus natural selection is expected to favour individuals with the behavioural mechanisms to make decisions as if they were computing Bayesian estimates in typically-experienced environments, although this does not necessarily imply that favoured decision-makers do perform Bayesian computations exactly. Each individual should evolve to behave as if updating a prior estimate of the unknown environment variable to a posterior estimate as it collects evidence. The prior estimate represents the decision-maker's default belief regarding the environment variable, i.e., the individual's default 'worldview' of the environment. This default belief has been hypothesised to be shaped by natural selection and represent the environment experienced by the individual's ancestors. We present an evolutionary model to explore how accurately Bayesian prior estimates can be encoded genetically and shaped by natural selection when decision-makers learn from uncertain information. The model simulates the evolution of a population of individuals that are required to estimate the probability of an event. Every individual has a prior estimate of this probability and collects noisy cues from the environment in order to update its prior belief to a Bayesian posterior estimate with the evidence gained. The prior is inherited and passed on to offspring. Fitness increases with the accuracy of the posterior estimates produced. Simulations show that prior estimates become accurate over evolutionary time. In addition to these 'Bayesian' individuals, we also introduce 'frequentist' individuals that do not use a prior and instead use frequentist inference when estimating the probability. Competition between the two shows that the former tend to have an evolutionary advantage over the latter, as predicted by the literature, and that this advantage is lowest when the information available to individuals poses the least uncertainty.

Keywords: Baldwin effect; Bayesian learning; Bayesian priors; Evolution; Natural selection; Optimal decision-making.