Fig. 2: A multi-start pipeline diversifies the optimisation start points.
From: Current state and open problems in universal differential equations for systems biology
Initial mechanistic parameters and hyperparameters are sampled repeatedly and used to initialise multiple optimisation runs. The contour plots indicate the objective function landscape, with the negative log-likelihood as the objective function. Note that the objective function landscape differs depending on the hyperparameter settings, and the model parameters are optimised in a gradient-based manner. Lower and upper bounds for a mechanistic parameter are indicated by dashed lines. Early stopping is used to determine the optimal parameter vector.
