Code and data for the bachelor's thesis "Dynamische Budget-Allokation in Multi-Objective Reinforcement Learning"

This record contains the code and the experimental data of the bachelor's thesis "Dynamische Budget-Allokation in Multi-Objective Reinforcement Learning" (Dynamic Budget Allocation in Multi-Objective Reinforcement Learning) by Jonas Reichle, submitted at the University of Augsburg, Chair of Organic Computing, on 28 September 2026. The thesis itself is not part of this record. Decomposition-based multi-objective reinforcement learning trains one policy per weight vector. The thesis studies how a fixed training budget should be distributed across these subproblems during training. Ten allocation heuristics, combined with three progress signals into 18 configurations, are compared under MOPPO and MOSAC on four MuJoCo tasks from MO-Gymnasium with seeds 1 to 10. The data comprise the raw results of the main matrix (1,440 runs), of the side experiments (520 runs) and of the follow-up sweep P3 (160 runs): training histories per checkpoint (CSV), Pareto fronts per checkpoint (NPZ), run configurations (JSON) and, as a separate package, TensorBoard logs. A further package holds all derived evaluation files in the state on which the thesis is based. The code snapshot contains the allocation layer, the modified framework of Kemper et al. and the complete evaluation chain; with it, the result tables and figures of the thesis can be regenerated within minutes. README.md describes the packages, the file formats, the mapping of configuration labels to the abbreviations of the thesis, and the reproduction steps.