Targeted Optimisation

Modified on Wed, 31 May 2023 at 07:43 AM

Description

Targeted Optimisation is a tool that, based on a trained model, finds the inputs to that model that generate the output value as close as possible to a user-defined set of output values.


Application

Often in engineering, regulations, customers or competitors enforce some performances to be met by a product. The main application of Targeted Optimisation will be to help engineers reach a product with the desired performances and characteristics much faster than through iterative approaches.


How to use

Step Configuration

When applying Targeted Optimisation, a model must already be trained beforehand.

  • Select Model: Select a trained model in the field Model.
  • Set Targets: Select how the target values should be defined.
    Enter values manuallySelect this option to enter target values for the outputs manually.
    Read values from tableSelect this option to read target values from an uploaded table.
    • Upload a tabular file with the target values to the File Manager and import into the notebook with the Tabular importer.
    • The column names in that file must match the name of the outputs columns used to train the model.
    • Required format of the table:
      Output1Output2Output3
      TargetValue1TargetValue2TargetValue3
    • Select the corresponding dataset in the field Table to read target values from (optional).
  • Set Contraints (Optional): If you want to restrict the input range in which the optimisation will be applied, the Restrict inputs box should be ticked. Sliders for all inputs will then appear and you can define the optimisation range for each input with these. 
    • An input can be restricted to a fixed value instead of a range by ticking the box Fixed.
    • If the range for an input is not changed it is not restricted. That is, the input range is only restricted if the sliders are changed from their initial range (or if the Fixed option is ticked).
  • When optimising the output, this option can be useful if certain inputs are more expensive, harder to change or even cannot be modified. Reminder: by limiting/fixing the inputs, the optimisation towards the target value will probably be less effective.

  • Algorithm Settings: Select an Optimisation method in this section. Two options are available:
    Iterative Search (Slow)This method uses differential evolution (stochastic) algorithm to find the best model inputs for optimising the fitness function. The iterative search involves starting with an initial solution and then systematically improving it by making small adjustments until the optimal solution is found. This process is repeated until the desired level of accuracy is achieved.
    The user can define:
    • Number of recommended designs, which will be the best designs found.
    • Limit number of iterations, which is the maximal number of repetitions the optimisation algorithm will run until it stops.
    Global Sampling (Fast)

    This method randomly samples inputs on the design region and calculates the output values. The output closer to the target value (based on the fitness function) is ranked higher on the recommended designs.

    The user can define:

    • Number of recommended designs, which will be the best designs among all explored designs.
    • Number of points to sample, which is the total number of random designs that will be generated and evaluated.

    The Global Sampling method is faster than the Iterative Search, however for models with a high number of inputs, the sampling can become sparse and an optimum value becomes less likely to be found.

  • Fitness function: The fitness function is used to evaluate how close the generated designs are to the target design. Thus, the algorithm aims to minimise the output of the fitness function, meaning that the generated designs are more similar to the target design. See below in the section “More on this step” for details.

Step Results

  • Recommended Designs: The step outputs a list of recommended designs ranked according to the fitness function, showing the input combinations and calculated output.

Visualisation of Step Results

  • Parallel Coordinates: Shows all the designs explored in a Parallel Coordinate plot, in grey and the recommended designs in red. This can be helpful in finding patterns in the recommended designs.
  • 2D Scatter Plot: Shows all the designs explored in a 2D point plot, in grey and the recommended designs in red. The axes of the plot can be selected in Visualisation Settings before the Targeted Optimisation is applied. This can be helpful when plotting two parameters against each other.

More on this step

Fitness functions

Below are the equations that describe the fitness functions available on the platform. In the following equations, the vector q contains the numerical output values of the generated design and p the numerical output values of the target design.

Euclidian distance (default)Read more on Euclidean distance.

Manhattan distance

Collinearity coefficient
Amplitude correlation coefficient
Custom …This option enables you to use another fitness function. There will be the option of writing any fitness function in Python. It is recommended that you use the default fitness function unless another fitness function has already been deemed better through further research.

Was this article helpful?

That’s Great!

Thank you for your feedback

Sorry! We couldn't be helpful

Thank you for your feedback

Let us know how can we improve this article!

Select atleast one of the reasons
CAPTCHA verification is required.

Feedback sent

We appreciate your effort and will try to fix the article