The ARLBench Performance Dataset

ARLBench is a benchmark designed for hyperparameter optimization (HPO) in Reinforcement Learning (RL). Given that we conducted several thousand runs to identify meaningful HPO test settings for RL, we have compiled these results into a dataset for future research and applications.

This dataset can be leveraged to:

• Meta-learn insights about the hyperparameter landscape in RL.
• Warm-start HPO tools by utilizing previously explored configurations.

Dataset Details

The dataset includes:

• Landscape data: 10 runs each for PPO, DQN, and SAC across:
  • Atari-5 environments
  • Four XLand gridworlds
  • Four Brax walkers
  • Five classic control environments
  • Two Box2D environments
• Optimization data: 3 runs per optimization algorithm for each algorithm-environment combination, covering:
  • Population-Based Training (PBT)
  • SMAC
  • SMAC with Multi-Fidelity
  • Random Search

Dataset Mapping

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

For optimization runs, it additionally includes:

• Optimization seed: Differentiates between the five optimization runs per algorithm-environment pair.
• Optimization step: Tracks configurations evaluated at different steps.

Example Usage

You can find example notebooks demonstrating how to use:

• Landscape data
• Optimization data

For more details, refer to the ARLBench paper.