What is SmartSelection?
It is well known that the efficiency of design optimization and approximation problems solution highly depends on proper algorithm and technique selection. Tedious tuning of optimization algorithm internal parameters or manual search of the proper approximation technique can consume a lot of time and do not lead to success.
SmartSelection is a technique in both design optimization and approximation tools in pSeven that automatically chooses the most efficient solution approach for a given type of problem and data. The principle of this technique is that the user should interact with the tool only in terms of the problem description and properties that are important and clear, not in terms of low-level details of the algorithms and techniques that user is unaware of or doesn't care about.
Set of hints and options in SmartSelection helps the user to describe the problem and desired solution from his point of view, not from the algorithmic point of view. It hides algorithms and techniques complexity so that the user could concentrate on the engineering problem itself. This opens expert level mathematics for design optimization and approximation even for non-math experts.
Provide problem description
SmartSelection for Optimization
Successful design optimization process highly depends on many technical aspects and may implement several optimization methods with different globalization and local algorithms. That’s why choosing of the static algorithm may not lead to the desired result.
SmartSelection for design optimization uses hints about the problem provided by the user for initial navigation and chooses the algorithm automatically, while its parameters are tuned adaptively during the solution process. In the case of severe performance degradation, it interrupts solution process and restarts with the next suitable solver.
Hints and options in SmartSelection for design optimization:
- Hints for input parameters: continuous/integer.
- Hints for constraints and objective functions: expensive/cheap evaluation, linear/quadratic/generic behavior.
- Option presets: analytical problem, smooth problem, noisy problem, heavily noisy problem, expensive problem.
- High-level options: stop criteria, globalization intensity, number of Pareto points.
Optimization with SmartSelection vs. open algorithms.
NSGA-II, Adaptive Scalarization, SmartSelection – 280 iterations each
SmartSelection for Approximation
Selection of static technique is often not enough to build an accurate approximation model. The best model type and parameters of approximation technique highly depend on a particular problem and given data. In other words, crucial information from the technical point of view might be not known beforehand.
SmartSelection for approximation grants automatic and adaptive technique choosing. For better approximation quality, different parts of the model can be built with different techniques.
Hints and options in SmartSelection for building approximation models:
- Hints about the data: linear/quadratic/discontinuous dependency, dependent outputs, tensor structure.
- Desired model properties: acceptable quality, smooth model, accuracy evaluation, exact fit, do not store training sample, NaN prediction.
- Approximation properties: validation type, internal validation, randomized training, training time limit.
Quality of approximation models built with SmartSelection vs. static techniques in pSeven