Fig. 19
From: Hierarchical intrinsically motivated agent planning behavior with dreaming in grid environments
Comparison of agents with different intrinsic modulation signals at Exhaustible resource task. Baseline—the agent without any intrinsic modulation. Agents with prefix “positive” have intrinsic signal from [0, 1] interval. Agents with prefix “negative”—from \([-1, 0]\) interval. The anomaly is simple prediction error for TM. Random is value from uniform distribution. Empowerment is the ideal four-step empowerment