snooker-testbed-phase4e-sac-v1

Phase-4e canary: SAC algorithm switch after 4 PPO attempts all produced identical std-drift trajectories (1.042 at step 2M). Job 7241287 on torch h200 gh105, 1h42m wall, ran 2026-04-26 17:11-18:53 UTC. n_envs=1, 500k total steps. RESULT: SAC produced a single spike to 2.25 at step 150k (FIRST run to beat random baseline 2.10!) but did not sustain it. Final score 0.00 at step 500k — policy DEGRADED across the run. Same foul-rate plateau (98-99%) as all PPO runs. Conclusion: the bottleneck is not the RL algorithm; it is structural — env action space (4D continuous), sparse reward landscape (98% foul → near-zero advantages), and exploration scale all conspire to make this task infeasible at the current sample budget regardless of PPO/SAC. The 2.25 spike shows the env CAN be solved transiently; just not stably.

Dataset Info

• Rows: 10
• Columns: 9

Columns

Column	Type	Description
step	Value('int64')	Global SAC timestep
curriculum_stage	Value('int64')	Stage at this step (stages 0/1 cleared early)
mean_score	Value('float64')	Mean over 16 eps. Spike to 2.25 at step 150k (FIRST run to beat random 2.10), then degraded. Final 0.00.
max_score	Value('float64')	Best episode score in eval (peak 12 at step 150k)
mean_shots	Value('float64')	Mean shots per episode (200 = truncated)
mean_efficiency	Value('float64')	mean_score / mean_shots
mean_foul_rate	Value('float64')	Stays 98-99% — same as all PPO runs
episodes	Value('int64')	16 eval episodes per eval
source_file	Value('string')	Original metrics_step_*.json

Generation Parameters

{
  "script_name": "sbatch/train_sim.sbatch (Phase 4e \u2014 SAC algorithm switch)",
  "model": "stable-baselines3 SAC, MlpPolicy net_arch=[256,256], ent_coef=auto, batch_size=256, buffer=200k",
  "description": "Phase-4e canary: SAC algorithm switch after 4 PPO attempts all produced identical std-drift trajectories (1.042 at step 2M). Job 7241287 on torch h200 gh105, 1h42m wall, ran 2026-04-26 17:11-18:53 UTC. n_envs=1, 500k total steps. RESULT: SAC produced a single spike to 2.25 at step 150k (FIRST run to beat random baseline 2.10!) but did not sustain it. Final score 0.00 at step 500k \u2014 policy DEGRADED across the run. Same foul-rate plateau (98-99%) as all PPO runs. Conclusion: the bottleneck is not the RL algorithm; it is structural \u2014 env action space (4D continuous), sparse reward landscape (98% foul \u2192 near-zero advantages), and exploration scale all conspire to make this task infeasible at the current sample budget regardless of PPO/SAC. The 2.25 spike shows the env CAN be solved transiently; just not stably.",
  "hyperparameters": {
    "algorithm": "SAC",
    "total_timesteps": 500000,
    "n_envs": 1,
    "batch_size": 256,
    "ent_coef": "auto",
    "buffer_size": 200000,
    "learning_starts": 1000,
    "tau": 0.005,
    "train_freq": 1,
    "gradient_steps": 1,
    "learning_rate": 0.0003,
    "gamma": 0.99,
    "curriculum_stages": [
      0,
      1,
      2,
      3,
      4
    ],
    "advancement_threshold": -400.0,
    "curriculum_min_episodes_per_stage": 100,
    "curriculum_window_size": 50,
    "reward_shot_cost": 0.01,
    "reward_pot_bonus": 0.5,
    "reward_completion_bonus": 10.0,
    "reward_position_shaping": 0.0,
    "reward_hit_any_ball_bonus": 2.0,
    "reward_foul_multiplier": 1.0,
    "code_version": "2026-04-24-phase3c",
    "commit": "1bd6da2"
  },
  "input_datasets": [],
  "experiment_name": "snooker-testbed",
  "job_id": "torch:7241287",
  "cluster": "torch",
  "artifact_status": "final",
  "canary": false
}

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-phase4e-sac-v1", split="train")
print(f"Loaded {len(dataset)} rows")

---