Fig. 3: MPL and Fleet outperform other models given large search budgets.
From: Symbolic metaprogram search improves learning efficiency and explains rule learning in humans
A Percentage of functions (100 total) acquired per model (subplots) by a given trial (11 total) with human median performance (n = 389 people; gray curve), 25%-75% human performance (dark gray band), and best-worst human performance (light gray band). We measure acquisition using the strict criterion of generating correct predictions on all future trials. B Ratio of model mean accuracy to human mean accuracy (n = 389 people) per concept (dots; 100 total) per model, with parity between models and humans (dotted line) and a kernel density estimate (colored regions). The crossbars show the median across functions with a 95% bootstrapped CI. Each model is associated with a unique color for easier comparison across figures.
