The version 2.11 brings another set of new features and internal improvements into NCS, in particular to:
- Accelerate NCS optimization workflows and extend their capabilities
- Improve data exploration, external compatibilities and team work around NCS experiments
Faster and More Powerful Optimization
1. Multi Parametrizer
NCS 2.11 introduces the possibility to combine multiple parametrizations in optimization runs to explore larger design spaces. You can now combine any number of parametrizations for a single optimization run and automatically discover some overall optimal designs across multiple design spaces. As NCS surrogate models are natively parameter-free, this allows you to go beyond fixed parametrizations at optimization time and merge design concepts -- such as morphing, parametric CAD or custom parametrizations.
2. Improved Data Manipulation in NCS Online
In this new version, the data manipulation in the online optimization experiments is made more convenient for the user, and faster at run time. In particular
- Caching of the training data allows you to gain some precious time during retraining when running long online optimization workloads (up to 5-10x speed-ups on the retraining phase).
- At the end of an online optimization run, you can export your data to create a new training dataset. This allows to use the online script not only as an optimization loop, but as a strategy to do an initial DoE where the data is optimally distributed to train surrogate models that have a high performance on high quality designs.
3. Optimization Task Interactive Reports
From 2.11, NCS offers task-level interactive report templates. In this version, we introduce such reports for optimization-related task types:
These reports allow to compare pareto fronts for predicted or simulated metrics across multiple experiments in the same task.
B. Easier Data Exploration and Team Work around NCS Experiments
4. Enhanced Data Preprocessing Report
The data preprocessing report was significantly improved to be more user-friendly while allowing for more robust outlier detection across various experiment types:
- Outlier detection now works by interactively defining upper and lower thresholds for each entry in the data
- These thresholds can then be used to filter outliers on-the-fly during online optimization loops.
5. Experiment Import/Export
It is now possible to export an experiment from a workspace, and import it into another workspace: this allows sharing your workflow with collaborators, including your configuration files and optionally your data and models as well. If exporting from a cloud workspace to another cloud workspace, the input data is transferred through the cloud for maximum performance.
6. Latin Hypercube Sampling for DoEs with NCS Parametrizers
NCS now offers different methods to perform DoEs using ncs generate. In particular, Latin Hypercube Sampling is now available to generate well distributed datasets. The same methods are also available for selecting the initial samples of optimization algorithms, to ensure you always start your optimization with a well balanced population.