Line Plot from Scientific Research

For each dataset D (with dataset sizes N {25, 50, 100, 200, 400}), we sample models from p ( m D ) using the Bayesian machine scientist 5 , select the MDL model (maximum p ( m D )) among those sampled, and use this model to make predictions on a test dataset {D}^{{\prime} } , generated exactly as D . We show the prediction root mean squared error (RMSE) of the MDL model on {D}^{{\prime} } as a function of N and s . For comparison, we also show the predictions from an artificial neural network (ANN, dotted lines; Methods). Since s is the irreducible error, predictions on the diagonal RMSE = s are optimal.

DOI