Probabilistic Inference for Bayesian Optimization (PI-BO)

PI-BO is based on the Bayesian optimization method and uses the probabilistic properties of the DIM-GP metamodel to select new designs based on the greatest potential for design improvements while taking model uncertainty into account. This is possible for both single- and multi-objective optimization applications, including Pareto optimization. It can also handle any number of constraints. The algorithm can be applied to high-dimensional optimization problems through the use of the Siembo extension, which automatically identifies the important parameter subspaces relevant for the optimization. It is also able to deal with noise and failed designs. This method is especially suitable for engineering problems where the evaluation of the designs goes along with high computational costs. It does not require gradients and is designed to find the global optimum even for non-convex and mutli-variate problems.

Node Settings

To access the options shown in the following table, double-click the Probabilistic Inference for Bayesian Optimization (PI-BO) system on the Scenery pane and switch to the Settings tab.

Options	Default Value	Description
Basic Settings
Maximum number of iterations	10	This value determines the number of calculations of the process chain, along with the iteration_size (`Maximum number of designs = num_start_designs + max_num_iter × iteration_size` unless other termination criteria is met.).
Number of design evaluations per iteration	1	Number of design evaluations carried out per iteration. This value determines the number of calculations of the process chain, along with the `max_num_iter`. If parallelism is not required, consider increasing the 'train_interval' instead.
Number of new start designs	10	Number of designs to be created in addition to the start designs for the initial model.
Stagnation iterations	0 (no stagnation)	Maximum number of allowed iterations without improving the objectives.
Run Mode	Maximal accuracy	High level configuration for setting up the PI-BO optimization algorithm with respect to application needs. Maximal accuracy - PI-BO prioritizes accuracy. It incorporates local optimization techniques and exploits all points created in an internal optimization step to provide the most precise results. Using this run mode, customers can expect high precission with the best possible optimization progress. However this run mode is approximately 30% slower compared to the Minimal computation time run mode. Minimal computation time - This run mode is dedicated to users who prioritize minimal computation time on the PI-BO algorithm. A global exploration strategy is used to identify an optimal candidate with minimal points per iteration (Efficient batches) in order to come up with the optimal design candidate in a short amount of time. While it excels in speed, there can be a slight compromise in accuracy. The loss of optimum solution provided by Maximal accuracy can range from 1% to 30%, depending on the complexity of the optimization problem.
Advanced Settings
Pareto optimal front	True (auto-detect)	By default, Pareto optimization is conducted for problems with multiple objectives. If disabled, the objective function will be the weighted sum of all objectives which is optimized like a single objective. The setting has no effect on single objective problems.
Internal optimizer iterations	500	The number of optimization iterations used for the internal optimization of the acquisition function.
Internal population size	25	The population size is used for the internal optimization of the acquisition function. It is recommended to increase it with increasing number of design parameters.
Local search	Enabled	Activating local optimization during the search for the next candidate. Local search could increase the computation time but improve the results.
Efficient batches	Disabled	Internally estimating the efficient iteration step size in case of multiple samples per iteration. This is a useful option for limited resources and it is possible to calculate more than a single sample at once.
DIM-GP Training Settings
Maximum epochs	500	Maximum number of training epochs used. Algorithm messages hint at the required number of epochs.
Batch size	0 (disabled)	Activates batch training, if the value is non-zero and smaller than the number of samples. Recommended at higher sample sizes.
Noisy data	False (disabled)	Activates batch training, if the value is non-zero and smaller than the number of samples. Recommended at higher sample sizes.
Execution options
Number of simultaneous trainings	1	Number of parallel trainings of objectives and constraints.
Job submit pattern (beta)	<jobscript> <arg1> <arg2> <arg3>	Job submit pattern for bash or command prompt for submitting jobs. The first argument `<arg1>` is for submitting jobs, the second argument `<arg2>` for defining submission pattern that allows a list of arguments to passed to the job, and the third argument `<arg3>` is for defining the directory. If last entry does not exist, action will be skipped and if empty, arguments will be submitted without an optional argument command.

Additional Options

To access the options shown in the following table, in any tab, click Show additional options.

Option	Description
Designs in parallel	Controls the resource usage of nodes in the system. When the check box is cleared (default), a value is chosen to ensure the best possible utilization of the child nodes. When the check box is selected, set the value manually to specify how many designs are sent to child nodes, limiting the maximum degree of parallelism for all children. Ansys recommends keeping the check box clear.
Auto-save behavior	Select one of the following options: No auto-save Actor execution finished Every n-th finished design (then select the number of designs from the text field) Iteration finished (for optimization, reliability) The project, including the database, is auto-saved (depending on defined interval) after calculating this node/system (either when the calculation succeeds or fails). By default, all parametric and algorithm systems have Every nth finished design 1 design(s) selected, all other nodes have No auto-save selected.

Option

Description

Designs in parallel

Controls the resource usage of nodes in the system.

When the check box is cleared (default), a value is chosen to ensure the best possible utilization of the child nodes. When the check box is selected, set the value manually to specify how many designs are sent to child nodes, limiting the maximum degree of parallelism for all children. Ansys recommends keeping the check box clear.

Auto-save behavior

Select one of the following options:

No auto-save
Actor execution finished
Every n-th finished design (then select the number of designs from the text field)
Iteration finished (for optimization, reliability)

The project, including the database, is auto-saved (depending on defined interval) after calculating this node/system (either when the calculation succeeds or fails).

By default, all parametric and algorithm systems have Every nth finished design 1 design(s) selected, all other nodes have No auto-save selected.