Cologne, July 8, 2025 – The THK.AI Forschungscluster at TH Köln is proud to announce the successful conclusion of the Data-Driven and Modeling (DDMO) Blackbox-Optimization Challenge 2025, an engaging project that put students‘ analytical and optimization skills to the test. Organized by Richard Schulz and Prof. Dr. Thomas Bartz-Beielstein from THK-AI, this challenge simulated real-world research conditions, requiring participants to optimize a complex black-box yield function under various constraints.
The core task for the students was to find the optimal combination of parameters to produce the highest yield through systematic experimentation. The challenge was designed to mirror practical scenarios, incorporating key elements such as working with a limited budget, dealing with measurement noise that changed across rounds, and developing efficient strategies to optimize parameters. Participants had to navigate a function with 15 parameters, discovering which of these internally affected the result. The experiment began on June 7, 2025, with new rounds starting automatically every Monday and Thursday, bringing 100 additional budget units each time. The cost of experiments varied based on the number of replications. Crucially, measurement noise varied, influencing the reliability of data, and participants were advised to conduct important experiments in rounds with low noise.
We extend our congratulations to the students who achieved the highest places:
1st Place: Mit Tushar Phanase
2nd Place: Sugin Sukumaran
3rd Place: Mohamed Raafat Mohamed Shattat
Mit Phanase, the first-place winner, shared insights into his successful approach as follows: „My methodology for solving this Noisy Agricultural Yield Problem is based on a combination of PCA-Augmented Gaussian Processes and Residual Modeling”.
Mohamed Shattat, who secured third place, explained his strategy: „I’ve trained surrogate models, specifically an ensemble of Gaussian Process Regressors combined with Gradient Boosting, on historical yield data. Using these models, I predicted outcomes for Sobol-sampled nutrient combinations and identified the optimal candidate via a Bayesian optimization acquisition function“.
The Blackbox-Optimization Challenge, part of the DDMO module – Summer Semester 2025, provided an invaluable learning experience, simulating a real-world agricultural scenario with constraints such as budget limits, measurement noise, and multiple rounds of experiments. This project not only facilitated practical application of optimization techniques but also encouraged students to develop their own adaptive strategies and creative approaches.