Fig. 14

Comparison of different dreaming switching strategies in four rooms with fixed positions experiments. Left: TD error-based switching strategy (green) does not add to performance of the baseline with no dreaming (red). Right: anomaly-based dreaming (red) shows a significant improvement over the baseline with no dreaming (green). It performs similarly to the baseline with the 50% increased learning rate (light blue) and converges twice faster than the baseline with the 25% reduced learning rate (blue, results are x2 shrunk along the X-axis)