ML2Grow houses over 25 years of expertise of a successful research group of Ghent University for optimization of complex systems. Our solutions are extremely powerful for the optimization of expensive problems. For these problems parameters such as the angle, geometry, material properties, production or environmental parameters are often not optimized because each new parameter combination involves a high cost: a new prototype needs to be built, a computationally demanding simulation needs to be performed or a test has to be performed in a lab or in the field in order to assess the performance or feasibility. Due to budget constraints, the number of tests that can be performed is restricted.
How do our solutions work?
Define the parameters: we reconfigure to operate on your problem domain.
Test and evaluate suggestions: our technology platform suggests values for the parameters. Use these suggestions to evaluate the performance of the problem.
Report performance: the performance is fed back into our system which calculates the next suggestion. Each iteration, this suggestion maximizes the probability of finding an optimal solution.
Repeat until optimal solutions are found: our solutions identify optimal solutions using 10 to 20 times less suggestions than traditional optimization techniques.
Wide variety of applications
Optimization: one or multiple objectives (i.e., identify and enhance trade-off between weight and size in an aerodynamical problem) with optionally uncertainty on the input parameters (expressing robustness of the solution to changing conditions).
Feasibility: identify parameter combinations for which certain specifications are met (or not met). This can be performed during an optimization problem itself, or performed after for root-cause analysis.
Human-in-the-loop: some knowledge is only available in the form of human expertise. This feedback can be included throughout the optimization process to guide the search for the next suggestion.
Multi-fidelity: during the evaluation of suggestions, it is possible to do this in several ways (with different levels of fidelity). This allows the use of, for instance, partially converged simulations to reduce costs.
A/B testing: in case only a comparison between several options is possible, we support tournament-based optimization.