Fig. 8
From: Tree-based learning for high-fidelity prediction of chaos
Sensitivity to hyperparameters in SOI training, where blue triangles are the overembedding dimension, k, orange squares are the successive time lags, \(\xi\), and green pentagons are the number of reduced features, p. Shaded regions indicate ±1 standard deviation with 1000 realizations of each point. Training and test data is equivalent to Fig. 6. (a) Root Mean Square Error (RMSE) and (b) Normalized Average Mutual Information between test and predicted SOI data. (c) Runtime of combined model training and testing in seconds. ‘Factor’ is the multiplicative factor applied to each hyperparameter, meaning 1.0 is simply the prescribed value of the given hyperparameter. The orange twin axis shows the value of \(\xi\) independently since \(\xi\) must be a natural number. Here, we take the default values of k and p as described earlier with an associated quantile threshold of 0.5, and the default value of \(\xi\) is 1.
