Fig 4. Illustration of the methodology and results of forward selection.

Black lines show results for the individual ‘AGBk’; other individuals are shown in gray. (a) states were added to the model by selecting the state providing the largest improvement when applied to the training set (dotted line). The supplemented model was then tested on the testing set (solid line), and the method repeated until all second-order states had been added to the model. The best (interpolated) model was that which gave the minimum negative log-likelihood when applied to the testing set (solid circles). (b) structure of resulting interpolated models, showing the proportion of a bird’s repertoire exhibiting zero-order, first-order and second-order properties.