Fig. 5: Recurrent agents achieve higher spatial information in partially observable environments.
From: Hippocampus supports multi-task reinforcement learning under partial observability
a Example place fields for hcDRQN and hcDQN neurons with the highest, median and lowest spatial information. b Spatial information (SI) distributions for all CA1 model neurons across hcDRQN and hcDQN for both the baseline environment (left) and the environment with a longer maze (right). Box plots show median (center line), interquartile range (box = Q1-Q3), and whiskers extending to the most extreme data within 1.5 × IQR from the hinges; points beyond the whiskers are considered outliers and are not displayed. Thus, the plotted min/max are the whisker endpoints, not the global extrema. Two-sided independent-samples t-tests on per-neuron SI values (n = 200). phcDRQN vs hcDQN = 1.95 × 10−2, phcDQN vs hcDRQN(full) = 2.38 × 10−4, phcDRQN(full) vs hcDQN(full) = 5.16 × 10−4, phcDRQN vs hcDRQN(full) = 1.22 × 10−8, phcDRQN vs hcDQN(full) = 4.56 × 10−4, phcDQN vs hcDQN(full) = 3.56 × 10−1. phcDRQN(long) vs hcDQN(long) = 6.45 × 10−4. *: p < 0.05, **: p < 0.01, ***: p < 0.001, ****: p < 0.0001, ns indicates no significant. Source data are provided as a Source Data file.
