Datasets:

aditijc
/

snooker-testbed-phase4g-obsaug-v1

step int64 50k 500k	curriculum_stage int64 2 2	mean_score float64 0.13 1.69	max_score float64 1 14	mean_shots float64 200 200	mean_efficiency float64 0 0.01	mean_foul_rate float64 0.98 0.99	episodes int64 16 16
50,000	2	0.75	8	200	0.00375	0.985313	16
100,000	2	1.25	10	200	0.00625	0.979063	16
150,000	2	0.25	1	200	0.00125	0.991562	16
200,000	2	0.125	1	200	0.000625	0.98625	16
250,000	2	0.375	5	200	0.001875	0.99	16
300,000	2	0.5	3	200	0.0025	0.98875	16
350,000	2	1.6875	10	200	0.008438	0.982188	16
400,000	2	1	14	200	0.005	0.986875	16
450,000	2	0.5625	7	200	0.002813	0.990312	16
500,000	2	1.0625	8	200	0.005312	0.988125	16

snooker-testbed-phase4g-obsaug-v1

Phase 4G: SAC + augmented 79-D observation. After all PPO/SAC runs hit a 98% foul plateau, hypothesis was that the policy couldn't discover snooker geometry from raw 73-D ball positions. Added 6 hand-crafted features: cue→best-target sin/cos, normalized cue→target & target→pocket distances, alignment score (cos angle between cue→target and target→pocket), has_target flag. Job 7250561 on torch h200, 1h57m wall, ran 2026-04-27 00:53-02:50 UTC. RESULT: AMBIGUOUS positive. Compared to Phase 4e (same SAC config, raw 73-D obs): final score 1.06 vs 0.00, mean 0.76 vs 0.64, variance 0.48σ vs 0.69σ. Less catastrophic forgetting, higher floor. But peak only 1.69 (< random 2.10), and foul rate plateau UNCHANGED at 98-99%. Verdict: obs aug helps stability but is NOT the lever to break the plateau. Proceeding to Phase 4F (discretize action space).

Dataset Info

Rows: 10
Columns: 8

Columns

Column	Type	Description
step	Value('int64')	Global SAC timestep
curriculum_stage	Value('int64')	No description provided
mean_score	Value('float64')	Mean over 16 eps. Peak 1.69 at step 350k. Less variance than 4e.
max_score	Value('float64')	Best ep score (peak 14 at step 400k)
mean_shots	Value('float64')	No description provided
mean_efficiency	Value('float64')	No description provided
mean_foul_rate	Value('float64')	98-99% (UNCHANGED from 4e — obs aug doesn't break the plateau)
episodes	Value('int64')	No description provided

Generation Parameters

{
  "script_name": "sbatch/train_sim.sbatch (Phase 4G \u2014 observation augmentation)",
  "model": "stable-baselines3 SAC, MlpPolicy net_arch=[256,256]",
  "description": "Phase 4G: SAC + augmented 79-D observation. After all PPO/SAC runs hit a 98% foul plateau, hypothesis was that the policy couldn't discover snooker geometry from raw 73-D ball positions. Added 6 hand-crafted features: cue\u2192best-target sin/cos, normalized cue\u2192target & target\u2192pocket distances, alignment score (cos angle between cue\u2192target and target\u2192pocket), has_target flag. Job 7250561 on torch h200, 1h57m wall, ran 2026-04-27 00:53-02:50 UTC. RESULT: AMBIGUOUS positive. Compared to Phase 4e (same SAC config, raw 73-D obs): final score 1.06 vs 0.00, mean 0.76 vs 0.64, variance 0.48\u03c3 vs 0.69\u03c3. Less catastrophic forgetting, higher floor. But peak only 1.69 (< random 2.10), and foul rate plateau UNCHANGED at 98-99%. Verdict: obs aug helps stability but is NOT the lever to break the plateau. Proceeding to Phase 4F (discretize action space).",
  "hyperparameters": {
    "algorithm": "SAC",
    "total_timesteps": 500000,
    "n_envs": 1,
    "obs_dim": 79,
    "obs_augmentation": "best-aim sin/cos, dist_target, dist_pocket, alignment, has_target",
    "code_version": "2026-04-26-phase4g-obsaug",
    "commit": "c5ba3db"
  },
  "input_datasets": [],
  "experiment_name": "snooker-testbed",
  "job_id": "torch:7250561",
  "cluster": "torch",
  "artifact_status": "final",
  "canary": true
}

Experiment Documentation

For complete experiment details, see https://github.com/aditijc/snooker-testbed

Usage

from datasets import load_dataset

dataset = load_dataset("aditijc/snooker-testbed-phase4g-obsaug-v1", split="train")
print(f"Loaded {len(dataset)} rows")

Downloads last month: 46

Number of rows:

Total file size:

4.84 MB