autorl-org/Risk-aware_ARLbench

The Risk-Aware ARLBench Performance Dataset

This dataset is with the paper Efficient Heteroscedastic Bayesian Optimization for Risk-Aware AutoRL, in which we propose efficient risk-averse Bayesian optimization (BO) method ERAHBO for reinforcement learning (RL) hyperparameter optimization, to be presented at Reinforcement Learning Conference (RLC) 2026.

The data set is an extension of ARLbench dataset, with a focus on reproducibility and reliability of reinforcement learning hyperparameters, in particular, it focuses on the variation of the policy return across random seeds.

Dataset details

The dataset includes:

• Landscape data: 50 runs each for PPO, DQN, and SAC across:
  • Four XLand gridworlds
  • Two Brax walkers
  • Five classic control environments

with 512 configurations generated by a random search.

Dataset Mapping

The dataset follows this mapping: $$\text{training steps, seed, hyperparameter configuration}\mapsto \text{training performance}\backslash text\{training\; steps,\; seed,\; hyperparameter\; configuration\}\; \backslash mapsto\; \backslash text\{training\; performance\}$$

Paper and code

• Paper (preprint): arXiv:2607.26680
• Code: github.com/LUH-AI/Efficient-Risk-Averse-BO

Citation

If you use this dataset, please cite:

@inproceedings{che2026erahbo,
  title         = {Efficient Heteroscedastic Bayesian Optimization for Risk-Aware AutoRL},
  author        = {Che, Mingxuan and Tseng, Tsung-Yuan and Eimer, Theresa and Lindauer, Marius and von Rohr, Alexander},
  booktitle     = {Reinforcement Learning Conference (RLC)},
  year          = {2026},
  note          = {To appear},
  eprint        = {2607.26680},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2607.26680}
}