Fig. 4: Assessment of the impact of data density and noise levels on the performance of UDE model of Glycolysis.
From: Current state and open problems in universal differential equations for systems biology
a Training loss for all data settings: The colour indicates the negative log-likelihood of the best UDE model by noise/sparsity setting, normalised by the number of data points used for training. b–d The number of fits in the 10 best fits (by training loss) per noise/sparsity setting that recover (in predictions) (b) oscillatory behaviour, (c) amplitude, or (d) frequency close to the true solution. (e) Best fits (= best training loss) from four different noise/sparsity settings: 16 (left column) and 61 (right column) data points per observable; and low noise (upper row, 5%) or high noise (lower row, 35%).
