Fig. 4: Generalization performance of the baseline agents.
From: Phy-Q as a measure for physical reasoning intelligence
a, Broad generalization of the baseline agents. b, Comparison of agents for local generalization and broad generalization (Phy-Q score) for nine agents: Bambirds (BB), Eagle’s Wing (EW), Datalab (DL), Pig Shooter (PS), D-DDQN-Symbolic (D-DDQN-Sy), D-DDQN-Image (D-DDQN-Im), Relational-Symbolic (Rel-Sy), Relational-Image (Rel-Im) and random (Random). The performance of the best-performing agent is shown in bold. Learning agents have higher local generalization values but lower values in broad generalization than the performance of heuristic agents. Human performance is way beyond agents.
