README.md

---
license: cc-by-4.0
pretty_name: "PPO Hyperparameter Corpus — From Importance Shifts to Landscape Topology (PPSN 2026)"
tags:
  - reinforcement-learning
  - PPO
  - hyperparameter-optimization
  - fitness-landscape-analysis
  - parallel-rl
  - automl
  - fanova
  - continuous-control
size_categories:
  - 10K<n<100K
task_categories:
  - tabular-regression
configs:
  - config_name: configs_with_hyperparameters
    data_files: data/configs_with_hyperparameters.csv
    default: true
  - config_name: all_runs
    data_files: data/all_runs.parquet
  - config_name: eval_trajectories
    data_files: data/eval_trajectories.parquet
---

# Data Release — From Importance Shifts to Landscape Topology: Characterizing and Exploiting Hyperparameter Spaces in Parallel RL

**Authors:** Yingjie Zou, Zhong Fan (University of Exeter, Exeter, UK)
**Venue:** PPSN 2026 (Parallel Problem Solving from Nature)
**License:** CC-BY-4.0

This HuggingFace dataset is the data and figure release accompanying the paper. It provides
the full per-run hyperparameter (HP) corpus, the validated analysis tables that back
every figure and table in the paper, the final paper figures, and the compiled appendix
(`supplement.pdf`).

## Corpus

The corpus consists of **37,200 PPO training runs** organised as:

- **5 chapters** (HP groups): Chap1 (PPO core), Chap2 (GAE), Chap3 (normalization),
  Chap4 (optimization), Chap5 (architecture)
- **× 4 Brax/MuJoCo continuous-control environments:** ant, halfcheetah, hopper, walker2d
- **× 5 parallelism levels:** `num_envs ∈ {64, 128, 256, 512, 1024}`
- **× 3 seeds:** {42, 43, 44}

Per-chapter configuration counts: Chap1/2/4/5 = 135 configs (8,100 runs each);
Chap3 = 80 configs (4,800 runs). Total = 37,200 runs. Every run completed
(`state == finished`) with a non-missing `final_eval_return`. The final performance
metric is `eval/episode_returns` (W&B summary), stored as `final_eval_return`.

Training used the `evorl` PPO implementation on NVIDIA GH200 (Grace Hopper) GPUs via
SLURM. The 37,200-run total reconciles exactly with the upstream training-run logs
(Chap1/2/4/5 = 8,100 runs each, Chap3 = 4,800).

## Package layout

```
04_huggingface/
├── README.md                              # this dataset card
├── .gitattributes                          # git-LFS tracking (*.parquet)
├── CHECKSUMS.sha256                        # sha256 of every released data/figure/document file
├── supplement.pdf                          # compiled appendix (supplementary material) PDF
├── data/
│   ├── configs_with_hyperparameters.csv   # headline per-run table (37,200 × 23)
│   ├── all_runs.parquet                   # full consolidated corpus (37,200 × 48)
│   ├── eval_trajectories.parquet          # eval-return trajectories (run × 10 snapshots)
│   └── derived/                           # validated analysis tables (back the paper)
└── figures/                               # final main + supplement paper figures (PDF)
```

## `configs_with_hyperparameters.csv` — column dictionary

One row per run; 37,200 rows × 23 columns. The `config.` prefix has been dropped from
HP names for readability; names follow `data/derived/chapter_hp_mapping.md`.

| Column | Type | Description |
|--------|------|-------------|
| `chapter` | str | HP group / chapter: `Chap1`–`Chap5` |
| `env_name` | str | Brax/MuJoCo env: `ant`, `halfcheetah`, `hopper`, `walker2d` |
| `num_envs` | int | Parallelism level (number of parallel envs): 64–1024 |
| `seed` | int | Random seed: 42, 43, 44 |
| `rollout_length` | int | PPO rollout length (steps per env per update) |
| `discount` | float | Discount factor γ |
| `gae_lambda` | float | GAE λ |
| `gae_horizon` | int | GAE horizon (0 = full rollout) |
| `clip_epsilon` | float | PPO clipping ε |
| `loss_weights.actor_loss` | float | Actor (policy) loss weight |
| `loss_weights.critic_loss` | float | Critic (value) loss weight |
| `loss_weights.actor_entropy` | float | Entropy bonus weight (negative = bonus) |
| `optimizer.lr` | float | Adam learning rate |
| `optimizer.grad_clip_norm` | float | Global gradient-norm clip (0 = off) |
| `normalize_obs` | bool | Observation normalization on/off |
| `normalize_gae` | bool | Advantage (GAE) normalization on/off |
| `minibatch_size` | int | SGD minibatch size |
| `reuse_rollout_epochs` | int | PPO epochs per rollout (sample reuse) |
| `agent_network.actor_hidden_layer_sizes` | str | Actor MLP hidden sizes, e.g. `[256, 256]` |
| `agent_network.critic_hidden_layer_sizes` | str | Critic MLP hidden sizes |
| `total_timesteps` | int | Training budget in env steps (fixed constant; see note below) |
| `final_eval_return` | float | **Outcome:** final evaluation episodic return |
| `eval_episode_lengths` | float | Final evaluation episode length (auxiliary) |

> **Note on `total_timesteps`:** this is a **fixed training-budget constant**
> (`10,485,760` ≈ 1×10^7) shared by every one of the 37,200 runs. It is therefore
> *not* a swept hyperparameter and is not part of the per-chapter swept-HP grid
> documented in `data/derived/chapter_hp_mapping.md`; it is retained in this table
> to fully specify each run's training budget.

Within each chapter, a subset of these HPs is swept and the rest are fixed; see
`data/derived/chapter_hp_mapping.md` for the per-chapter swept/fixed split and grids.

## File → paper artifact mapping

| Released file | Backs |
|---------------|-------|
| `data/configs_with_hyperparameters.csv` | Headline per-run corpus; input to all fANOVA / landscape / HPO analyses |
| `data/all_runs.parquet` | Full consolidated corpus (48 cols incl. training/eval metrics) |
| `data/eval_trajectories.parquet` | Per-run eval-return trajectories (10 snapshots/run); stage-wise analysis input |
| `data/derived/fanova_individual_importance.csv` | Table 3; Fig S1 (`figS_A1_fanova_heatmap`), Fig S4 (`figS_A4_importance_shift_ratio`) — HP importance shifts |
| `data/derived/fanova_pairwise_importance.csv` | Pairwise fANOVA importance (HP interaction analysis) |
| `data/derived/stagewise_importance.csv` | §4.2 stage-wise importance; Fig S3 (`figS_A3_stagewise_evolution`) |
| `data/derived/graphfla_landscape_features_validated.csv` | §5 landscape; Table 4; Fig S5 (`figS_A5_graphfla_feature_trends`), Fig S6 (`figS_A6_landscape_profile`), Fig S8 (`figS_A8_feature_method_corr`) |
| `data/derived/hpo_results_v2.csv` | §6 HPO benchmark; Fig 3 (`fig_main_results`); Fig S7 (`figS_A7_hpo_benchmark`), Fig S8 |
| `data/derived/selector_n100_perf.csv` | §S5 selector study (n=100 per-seed performance); Fig S11 (`figS_A14_selector_n100_accuracy`) |
| `data/derived/selector_n100_robustness.json` | §S5 selector robustness (10 optimiser-seed headline numbers); Fig S11 |
| `data/derived/robust_trends_R1-8.csv` | Table S2; block-structure-robust trend tests; ρ* labels in Fig S5 |
| `data/derived/mayor_metrics_v2_all.csv` | §S6 network diagnostics; Fig S12 (`figS_mayor_metrics`) |
| `data/derived/seed_reliability.csv` | Seed-reliability validation (cross-seed agreement) |
| `data/derived/discretization_validation.csv` | Grid-discretization validation |
| `data/derived/chapter_hp_mapping.md` | Per-chapter swept/fixed HP dictionary and grids |
| `figures/fig_main_results.pdf` | Main: HPO benchmark results |
| `figures/fig_landscape_trends.pdf` | Main: landscape-feature trends vs parallelism |
| `figures/fig_stagewise.pdf` | Main: stage-wise importance |
| `figures/fig_importance_shift.pdf` | Main: importance shift vs parallelism |
| `figures/figS_*.pdf` | Supplement figures (S1–S12); see `MANIFEST` notes in the supplement |
| `supplement.pdf` | Compiled supplementary material (appendix): full text of §S1–S6 and Figures S1–S12 |

Note on figure numbering: supplement figure files are named by their internal label
(`figS_A1`…`figS_A14`, `figS_mayor_metrics`); the paper text refers to them as
Figures S1–S12 in order. The mapping above gives both for the key analysis figures.

> §S6 (network diagnostics) was corrected for the camera-ready to use
> `mayor_metrics_v2_all.csv`. The earlier per-step plasticity CSVs are **superseded**
> and are intentionally **not** part of this release.

## Regenerating the figures

- **Main figures** (`fig_main_results`, `fig_landscape_trends`, `fig_stagewise`,
  `fig_importance_shift`): `EXP/analysis_scripts/build_main_figures.py`
- **Supplement figures** (`figS_*`): `supplement/_make_supplement_figs.py`

Both scripts read the derived CSVs released here (and the consolidated parquet) and
emit the vector PDFs in `figures/`. The analysis tables themselves are produced by the
scripts in `EXP/analysis_scripts/` (`run_fanova.py`, `run_graphfla.py`,
`run_hpo_benchmark.py`, `run_stagewise_fanova.py`, `run_selector_n100.py`,
`compute_robust_trends.py`, `finalize_validated_graphfla.py`) from
`all_runs.parquet` / `configs_with_hyperparameters.csv`. These analysis/figure scripts
are part of the paper's code repository and are not bundled in this data release.

## Integrity

- `CHECKSUMS.sha256` lists the sha256 of every released data, figure, and document file.
  `eval_trajectories.parquet` sha256 =
  `e7ef8116687e4bee178f04800c649ac35b03bc1806fe66cd5b175f5f2b70cb1f`
  (matches the project `DATA_PROVENANCE.md`).

## Reproducibility & data availability

All derived quantities in the paper and supplement are computed from the raw corpus
(`data/all_runs.parquet`, `data/eval_trajectories.parquet`; 37,200 PPO runs across
5 HP chapters, 4 Brax environments, 5 parallelism levels, and 3 seeds) using the
analysis scripts provided with the paper. Per-file source hashes are in
`CHECKSUMS.sha256`; full regeneration paths are recorded in the project's
`DATA_PROVENANCE.md`.

## Citation

Please cite the PPSN 2026 paper. License: CC-BY-4.0.