8 Proposing new points

A video presentation of this section can be found here.

To generate a set of non-implausible points, based on the trained emulators, we use the function generate_new_design, exactly as in the deterministic case:

new_points <- generate_new_design(stoch_emulators, 150, targets)

Here we generated \(150\) points, since we are going to use \(100\) of them to train new emulators and \(50\) of them to validate new emulators (as done in the first wave). We can visualise the non-implausible space at the end of this first wave using plot_wrap:

plot_wrap(new_points, ranges)

Here we see which parameters are more or less constrained at the end of the first wave. For example, it seems clear that low values of \(\gamma\) cannot produce a match (cf. penultimate column). We can also deduce relationships between parameters: \(\beta_1\) and \(\epsilon\) are an example of negatively-correlated parameters. If \(\beta_1\) is large then \(\epsilon\) needs to be small, and vice versa. Other parameters, such as \(\omega\) or \(\mu\), are instead still spread out across their initial range.

As shown in Tutorial 2, it is also possible to perform a full wave of emulation and history matching using the function full_wave, which needs the following information:

  • A dataset that will be split by the function into training data and test data;

  • A list of ranges for the parameters;

  • The targets: for each of the model outputs to emulate, we need a pair (val, sigma) or (min, max) that will be used to evaluate implausibility.