snooker-testbed-phase4f-discrete-v1

Phase 4F BREAKTHROUGH RUN. After 13 prior runs all hit a 98% foul-rate plateau, discretizing the action space (continuous Box[-1,1]^4 → MultiDiscrete([36 phi, 8 force, 5 a-spin, 5 b-spin]) = 7200 actions) finally broke through. Job 7308075 on torch h200, 19m53s wall, ran 2026-04-27 13:16-13:36 UTC. RESULTS (vs prior best run Phase 4G with augmented obs only): - mean over 10 evals: 1.25 (vs 4G 0.76, 4e 0.64) — 2× improvement - peak score: 3.50 at step 300k (vs 4G 1.69, 4e 2.25 single spike) - max single-ep score: 14 (vs 4G's 14, 4e's 12) - foul rate broke 97% TWICE (96.8% at step 300k, 96.4% at step 450k) — FIRST time below 97% in any run - score > 2.0 at TWO evals (3.50, 2.81) — sustained performance, not a spike Hypothesis confirmed: 4D continuous Gaussian was the bottleneck. Categorical PPO head directly upweights bins that score well, skipping the Gaussian-fit-over-chaotic-foul-space problem. Augmented obs (from Phase 4G) preserved in the recipe.

Dataset Info

• Rows: 10
• Columns: 8

Columns

Column	Type	Description
step	Value('int64')	Global PPO timestep
curriculum_stage	Value('int64')	No description provided
mean_score	Value('float64')	Mean over 16 eps. Peaks 3.50 at step 300k AND 2.81 at step 450k — sustained > 2.0 performance, first run to do so.
max_score	Value('float64')	Best single-ep score (14 at step 300k)
mean_shots	Value('float64')	No description provided
mean_efficiency	Value('float64')	No description provided
mean_foul_rate	Value('float64')	Hit 96.4% at step 450k — FIRST run to break below 97%.
episodes	Value('int64')	No description provided

Generation Parameters

{
  "script_name": "sbatch/train_sim.sbatch (Phase 4F \u2014 discrete action space)",
  "model": "stable-baselines3 PPO, MlpPolicy net_arch=[256,256], MultiDiscrete([36,8,5,5])",
  "description": "Phase 4F BREAKTHROUGH RUN. After 13 prior runs all hit a 98% foul-rate plateau, discretizing the action space (continuous Box[-1,1]^4 \u2192 MultiDiscrete([36 phi, 8 force, 5 a-spin, 5 b-spin]) = 7200 actions) finally broke through. Job 7308075 on torch h200, 19m53s wall, ran 2026-04-27 13:16-13:36 UTC. RESULTS (vs prior best run Phase 4G with augmented obs only):   - mean over 10 evals: 1.25 (vs 4G 0.76, 4e 0.64) \u2014 2\u00d7 improvement  - peak score: 3.50 at step 300k (vs 4G 1.69, 4e 2.25 single spike)  - max single-ep score: 14 (vs 4G's 14, 4e's 12)  - foul rate broke 97% TWICE (96.8% at step 300k, 96.4% at step 450k)    \u2014 FIRST time below 97% in any run  - score > 2.0 at TWO evals (3.50, 2.81) \u2014 sustained performance, not a spike Hypothesis confirmed: 4D continuous Gaussian was the bottleneck. Categorical PPO head directly upweights bins that score well, skipping the Gaussian-fit-over-chaotic-foul-space problem. Augmented obs (from Phase 4G) preserved in the recipe.",
  "hyperparameters": {
    "algorithm": "PPO",
    "total_timesteps": 500000,
    "n_envs": 8,
    "n_steps": 512,
    "batch_size": 2048,
    "ent_coef": 0.01,
    "learning_rate": 0.0003,
    "gamma": 0.99,
    "discrete_actions": true,
    "discrete_phi_bins": 36,
    "discrete_force_bins": 8,
    "discrete_spin_bins": 5,
    "obs_dim": 79,
    "obs_augmentation": "best-aim sin/cos, dist_target, dist_pocket, alignment, has_target",
    "code_version": "2026-04-27-phase4f-discrete",
    "commit": "76cf578"
  },
  "input_datasets": [],
  "experiment_name": "snooker-testbed",
  "job_id": "torch:7308075",
  "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-phase4f-discrete-v1", split="train")
print(f"Loaded {len(dataset)} rows")

---