Roman Küble
PhD Researcher (Organic Computing), University of Augsburg
I'm doing research in reinforcement learning and scene graph-based visual reasoning—with a focus on robust, interpretable models and reproducible results.
Skills & Technologies
Selected Projects
Master's Thesis
Learning to Explore: Reinforcement Learning with Scene Graph Integration for Full Environment Coverage
Modernised an ESSGG agent inside Ai2-THOR by replacing the RNN/REINFORCE baseline with a Transformer for long-horizon context and an Advantage Actor-Critic controller; experiments reveal that the stable A2C learner drives the gains in scene-graph quality, exploration efficiency, and generalisation, while the Transformer performs on par with the original RNN.
Full Paper
Evaluating Adaptive Systems: A Comparative Study of XCS and Established Reinforcement Learning Methods in Noisy Multi-Step Environments
Co-authored an empirical comparison of XCS and established reinforcement learning algorithms across 51 tasks from five benchmarks, demonstrating that RL remains markedly more reliable in complex multi-step settings while XCS only keeps pace on simpler, noisy problems (accepted, not yet published).
Vision Paper
Adaptive by Design: Rethinking the MLOC Architecture for Learning Systems
Helped redesign MLOC with lightweight RL agents in Layer 1 and a retunable Layer 2 simulator that an anomaly detector keeps in sync by separating model drift from agent uncertainty, enabling safe adaptation in dynamic environments (accepted, not yet published).
Code Contribution
MeshGraphNets.jl
Extended the Julia port of MeshGraphNets with native graph objects, custom plotting utilities, and a Pluto.jl notebook that showcases the new workflow.
Contact
Interested in collaborating or chatting? I’d love to hear from you.