Solvers
In this section, we describe first the general solver strategy, how the chemical network equations are solved. Followed by the available optimizers.
General solver strategy
The chemical network equations are solved by using numerical optimization methods. Currently the AdamW and the AdamaxW algorithms are implemented. Different optimizers could be added.
Using multiple walkers
For most initial conditions the optimizer is able to find the global solution within 100’000 to 200’000 steps. But sometimes, the solver can find only a local minimum. Since the solver takes only a fraction of a second to run, it is most of the time sufficient to use multiple walkers to avoid being stuck in a local minimum. As default we use 100 walkers.
Using multiple iterations
To further improve the results, multiple iterations can be used. For that the best nBestSolutions are used to generate a new set of walkers, by add some perturbation to the best results.
Optimizers
In the following all available optimizers are described. More optimizers could be added to the code.
AdamW
An iteration of the AdamW optimizer includes these steps:
with
\(\eta\): learning rate, user parameter (default = 0.4).
\(\lambda_{reg}\): Weight decay, user parameter (default = 0.0001).
\(\beta_1 = 0.9\)
\(\beta_2 = 0.999\)
\(\epsilon = 10^{-6}\)
AdamaxW
An single iteration of the AdamW optimizer includes these steps:
with
\(\eta\): learning rate, user parameter (default = 0.4).
\(\lambda_{reg}\): Weight decay, user parameter (default = 0.0001).
\(\beta_1 = 0.9\)
\(\beta_2 = 0.999\)
\(\epsilon = 10^{-6}\)