November 6, 2019
pSeven 6.16 Release
DATADVANCE development team announces the availability of pSeven 6.16, a new version of our flagship platform for data analysis and optimization. This is a major release bringing notable enhancements to design exploration and predictive modeling tools. It also contains several new features which simplify certain routine operations, and adds a number of usability and interface improvements. Read on for more information on the release highlights.
Speaking of design exploration, the most significant update in pSeven 6.16 is the new optional local search algorithm added to the Surrogate-based optimization (SBO) technique in the Design space exploration block (enabled with the technique's Local search option). This is an adaptive algorithm, which intelligently limits the search scope in optimization, adjusting the SBO technique's complexity to the response evaluation time. Local search enables effective surrogate-based optimization in problems with a high number of variables and responses. In many tasks, it also reduces the overall time in solving due to better balancing between the internal analysis and response evaluation stages.
A few more notable updates to the Design space exploration block are:
- You can now configure the Surrogate-based optimization technique to extend the initial sample with just one design (next best candidate) by limiting the number of response evaluations to 1.
- The Adaptive design and Latin hypercube sampling techniques now by default perform design space normalization internally, which makes their behavior more stable.
- The updated Adaptive design technique better handles the cases where response evaluation failures can occur.
Additionally, there are the following useful updates to the Uncertainty quantification block:
- Added the ports to output PDF and CDF data. You can use this data to create PDF and CDF plots in Analyze.
- The block can now save the analysis report to a HTML file on disk, in addition to the report which it sends to an output port. You can view the report saved to file, or use a Page viewer in Analyze to load the report received from the block's output.
- Added the multinomial distribution support, which enables user-defined discrete distributions.
Predictive modeling tools also received important updates in pSeven 6.16. First of all, the following internal improvements in SmartSelection algorithms further increase model accuracy and training performance:
- SmartSelection uses a new model validation procedure when estimating quality of intermediate models in the case when you do not supply a separate test sample. The new method is generally faster and provides more accurate quality estimates, eventually resulting in more accurate final models.
- The set of the Gaussian Processes (GP) technique options tuned by SmartSelection was extended, potentially allowing it to create more accurate GP models.
- When configuring SmartSelection, you now have an option to remove cross-validation data samples from the model. This potentially reduces the model size, hence its memory consumption, at the cost of removing some raw validation data.
Next one, SmartSelection is now supported by the ApproxBuilder block. In the SmartSelection training mode, the block automatically selects and configures an approximation technique to obtain the most accurate model. This mode is now default, and you can switch back to the manual configuration in the block's dialog.
Another new feature in model training is the support for constrained model input domain. You can add input constraints to a model, so it fits the training sample data but returns NaN outputs when input values are out of domain (do not satisfy the domain's constraints). This feature is supported both by the ApproxBuilder block and predictive modeling tools in Analyze. The input and output constraints are also shown in model details.
The Approximation model block now outputs model information and validation data as a report in the HTML format, which you can view in Analyze or export to a file on disk.
Finally, since 6.16 Approximation model can load and evaluate data fusion models trained by the DFBuilder block. In particular, this enables data fusion model export to all formats supported by the block, including various source code formats.
pSeven 6.16 adds some functions aimed to simplify routine post-processing and data analysis operations:
- Composite blocks can cache their input and output data to a file on disk, although it is often hard to analyze this data because cache files use a specific format. This is no longer a problem because you can import the cached data to a report and process it with any tool available in Analyze. For example, load it to a Sample viewer, or export to Excel or CSV.
- You can copy correlation analysis and variable importance data from a Sample viewer.
- The data import tool shows increase in performance when working with large Excel files. It also provides better table header recognition, and contains a few more usability tweaks.
Usability and Interface
This release adds various convenience features and improvements regarding interface responsiveness.
- Copy and paste values of inputs, outputs and parameters in Run. For example, use data from Excel or a text file to configure your workflow, or copy some workflow's outputs to another workflow's inputs.
- Copy values from a Sample viewer and paste them to the Inputs or Parameters pane in Run to set workflow's inputs or parameters.
- Test a script in the PythonScript block before running the workflow.
- Switch any Design space exploration block variable to constant directly from Run, without opening the block's configuration dialog.
- Edit names of multiple data series in the report database at the same time – just use the Rename... command with two or more data series selected.
- Sample viewer became more responsive as there is no unnecessary background data processing anymore – for example, when you switch to the Dependency tab or change settings there.
- Model validator no longer reloads data when you hide or show models, also avoiding unnecessary slowdowns.
- 2D plots which contain much data or many datasets are drawn noticeably faster.
pSeven 6.16 also includes numerous smaller improvements and bugfixes – please see the release changelog for a full list. You can also contact us to get more information and pSeven updates.