Learning performance dependence on model design: structure learning and imputation rule choice.

<p>(A) Each model design was evaluated with both active and random learners on two simulated 100 target x 100 condition datasets, each having eight phenotypes, 80% responsiveness and 40% uniqueness. For each model design the best average accuracy for either the active or random learner is plotted at each batch. For six cases displaying superlinear performance, structure learning methods are indicated in color, with different design variations plotted as separate lines and with filled circles to indicate batches where the active learner had higher accuracy: Greedy Merge (blue), a ‘strict’ variation of Greedy Merge (red), and B-Clustering (green, one design). These each had both Target Equivalence Class and Three-Point Imputation rules. (B) The difference in random and active learner accuracies for the superlinear model designs with structure learning method plotted by color as above; filled circles at tails indicate that the active learner had reached 100% accuracy.