Marvin Gülhan

Focus: AI & Algorithms · Current grade: 1.1 (1.0 = best)

Focus: Algorithms · Graduated with honors · Grade: 1.2 (1.0 = best)

Thesis (grade: 1.0): adapted Hybrid Genetic Search (PyVRP) to the Dynamic VRPTW, outperforming three state-of-the-art ACS baselines on the majority of dynamic Solomon instances.

Grade: 1.0 (1.0 = best)

Online combinatorial optimization for dynamic routing and resource allocation: contracted Germany-wide OpenStreetMap road graphs with more than 10M edges for use in disaster-logistics optimization and developed heuristic baselines for GNN-based reinforcement learning policies, including for the Airlift Planning Problem.

Designed and optimized the rear-wing aerodynamic package for the Formula Student team's 2024 race car through iterative CFD simulations using OpenFOAM.

Introduction to aerospace engineering and research methods, including clustering methods for anomaly detection in the fuel combustion of hybrid engines.

Shows that reinforcement learning (GRPO after a brief SFT warmup) can induce broad misalignment in a language model even under seemingly harmless reward signals, such as poor rhetoric or unpopular aesthetic preferences, and evaluates several mitigation techniques.

Presented at the 2026 International Conference on Military Communications and Information Systems (ICMCIS) in Bath, UK: a graph reinforcement learning approach using GNNs and MAPPO for complex dynamic planning problems such as the Airlift Planning Problem.

Designed the core routing approach for heterogeneous aircraft under capacity constraints and dynamic disruptions, contributing to a 2nd-place finish among 40 teams in the Air Force Research Lab (AFRL) Airlift Challenge.

Merit-based scholarship for outstanding academic performance.