Question about Serial scheduler and parallel scheduler

[zoechoutw]

Thank you for the amazing work! I have questions about the definition.

1. I wonder if @Serial scheduler is to execute the original Bayesian optimization algorithm without parallelization?
2. Could I assume that if I don't specify the scheduler, Serial scheduler is used by default?
3. To perform Bayesian optimization in parallel, what are the definition of n_jobs and batch_size for? What happens if the numbers are different? ex. n_jobs=m and batch_size=n. Does it run m runs or n runs simultaneously?

Thank you very much!

[tihom]

@zoechoutw thanks for the questions. Here are the answers:

1. I wonder if @Serial scheduler is to execute the original Bayesian optimization algorithm without parallelization?

Yes, it runs with batch_size=1 and evaluates the objective function one sample at a time.

2. Could I assume that if I don't specify the scheduler, Serial scheduler is used by default?

No. In that case, the batch_size=1 by default but the user has to modify the objective function to accept a list of params and return a list of results. See the notebook here for a working example.

3. To perform Bayesian optimization in parallel, what are the definition of n_jobs and batch_size for? What happens if the numbers are different? ex. n_jobs=m and batch_size=n. Does it run m runs or n runs simultaneously?

The batch_size is overwritten with the value provided for n_jobs when @scheduler.parallel(n_jobs=xx) is used.

[zoechoutw]

Thank you for your clear explaination!