Extended Data Fig. 4: Effect of the noise level on the convergence speed of the Morris water maze navigation task.
From: Actor–critic networks with analogue memristors mimicking reward-based learning
(a) Extracted update noise for both potentiation and depression curves of all in-software-emulated memristors. A single parameter σ, corresponding to the standard deviation of all 10 overlapped potentiation or depression measurement curves was extracted for each measured device. (b) Number of steps per episode plotted against the number of episodes for different noise levels, compared to the memristor emulation in the paper. The curves show the mean of 100 distinct simulation runs with an applied running average of 10 episodes to improve comparability between the runs. The in-software-emulated memristors exhibit noise levels between 2.9% and 10%, with a mean of 6.7%.
