Number of successful homing episodes for all five selected agents from each agent architecture, across different plume configurations for the same set of 240 initial conditions across varying agent starting location and head direction, and plume simulator state. ‘MLP_ X ’ refers to feedforward networks with X time steps of sensory history. Across all plume configurations, RNNs generally outperform feedforward networks, with more pronounced gains for more complex, switching wind direction (‘switch-once’, ‘switch-many’) plume tasks. In feedforward networks, performance on plumes with switching wind direction can improve statistically significantly with increasing memory. However, no statistically significant effect was observed for plumes with constant wind direction. Regression lines (solid black) are fitted on only MLP data ( N = 30, five agents per MLP type), but are extended slightly (dotted line) for comparison with RNNs ( P values are for a two-sided Wald test with the null hypothesis that the slope is zero).
