Fig. 1.
Structure learning improves prediction accuracy. (A) With structure learning. A simulated agent’s posterior probability over the upcoming decay rate on each planet is plotted. If the forager’s prior allows for the possibility of multiple clusters (α > 0), they learn with experience the cluster-unique decay rates. Initially, the forager is highly uncertain of their predictions. However, with more visitations to different planets, the agent makes increasingly accurate and precise predictions. (B) Without structure learning. If the forager’s prior assumes a single cluster (α = 0), the forager makes inaccurate and imprecise predictions—either over or underestimating the upcoming decay, depending on the planet type. This inaccuracy persists even with experience because of the strong initial assumption. Uncertainty adaptive discounting. (C) The effect of γcoef. The entropy of the posterior distribution over patch type assignment is taken as the forager’s internal uncertainty and is used to adjust their discounting rate, γeffective. The direction and magnitude of uncertainty’s influence on the discounting rate are determined by the parameter, γcoef. The more positive the parameter is, the more the discounting rate is reduced with increasing uncertainty, formalized as entropy. If negative, the discounting rate increases with greater uncertainty. (D) The effect of γeffective on overharvesting. Increasing γbase increases the baseline discounting rate, while increasing the slope term increases the extent the discounting rate adapts in response to uncertainty. (E) Overharvesting increases with α and γcoef in single patch type environments. Simulating the model in multiple single patch type environments with varying richness, we find that increasing α and γcoef, holding γbase constant, increases the extent of overharvesting (PRT relative to MVT). The richness of the environment determines the extent of the parameters’ influence, with it being greatest in the poor environment.