File size: 1,076 Bytes
9079e7b bc6cbc1 d1456b0 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 |
---
license: bsd-3-clause
tags:
- autorl
- automl
- rl
pretty_name: ARLBench Performance Data
size_categories:
- 10K<n<100K
configs:
- config_name: ppo_data
data_files:
- "atari_battle_zone_ppo.csv"
- "atari_double_dunk_ppo.csv"
- config_name: dqn_data
data_files:
- "atari_battle_zone_dqn.csv"
- "atari_double_dunk_dqn.csv"
- config_name: sac_data
data_files: "box2d_bipedal_walker_sac.csv"
---
# The ARLBench Performance Dataset
[ARLBench](https://github.com/automl/arlbench) is a benchmark for hyperparameter optimization in Reinforcement Learning.
Since we performed several thousand runs on the benchmark to find meaningful HPO test settings in RL, we collect them in this dataset for future use.
These runs could be used to meta-learn information about the hyperparameter landscape or warmstart HPO tools.
In detail, it contains each 10 runs for PPO, DQN and SAC respectively on the Atari-5 environments, four XLand gridworlds, four MuJoCo walkers, five classic control and two Box2D environments.
For more information, refer to the ARLBench paper. |