6.2.1. Optimization in ParAMS 2023.1

Starting with ParAMS 2023.1, the parameter optimization is controlled by GloMPO: Globally Managed Parallel Optimization.

To run a parameter optimization, set Task Optimization in the input file.

You can now run multiple optimizers in parallel. Optimizers converging to bad solutions or otherwise performing badly can be shut off and new ones started. You may configure:

• One or multiple Optimizers

• Exit conditions, see also the table below.

• Conditions for stopping optimizers, see also the table below.

• Parallelization options

• Output to generate

• Checkpoints (restart/continue from a previous run)

• The initial starting parameters for each new optimization (Generator)

• Parameter constraints

• Other input

	ExitConditions (link)	Stoppers (link)
What does it do?	Stops the entire optimization.	Stops one running optimizer.
What happens once triggered?	The job ends.	Other optimizers are unaffected and keep running. If possible, a new optimizer will be started to replace the stopped one.
When is it evaluated?	Approximately every 10 function evaluations or more frequently.	This is controlled by the CheckStopperInterval key.
Am I required to use one?	Yes	No
What is checked?	Generally, global properties like total time used, or total function calls made by all optimizers.	Generally, optimizer specific properties. Comparisons are often made between one optimizer and the best optimizer.
Easy to setup?	Yes. Generally determined by resource limitations (like time), or specific goals (like getting a certain number of results.	Generally, trickier to setup. Requires problem-specific information and a good guess of how optimizers will behave on the problem.
Good to know?	None.	By default, wont trigger for the best optimizer unless ApplyStoppersToBestOptimizer Yes is used.

The default behavior is to just run a single optimization, as in ParAMS 2022.1. This corresponds to the following input:

ParallelLevels
    Optimizations 1
End
ExitCondition
    Type MaxOptimizersConverged
    MaxOptimizersConverged 1
End