Spaces:
Sleeping
Sleeping
Commit History
-
v0.6.0: make height_breach_penalty configurable per TetrisEnv instance
Allows Phase 1 (5 pieces) to disable height breach while Phase 2 enables it. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>86f507b -
v0.5.1: re-enable decaying height breach penalty from line 4
Height breach penalty back at -50/level, decaying by 5 per piece locked. Combined with 5-piece curriculum, teaches model not to stack high early. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>2038b66 -
v0.5.0: reduce game_over to -50, disable height breach penalty
Reward signal was dominated by constant penalties (-500 game over + ~-350 height breach), leaving only ~5% learnable signal. Now learnable components are ~80% of total reward. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>3a5b76e -
chore: add version string to game_engine.py (0.4.0)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>cd64402 -
feat: decaying height breach penalty based on pieces placed
Height breach penalty now decays by 5 per piece locked: piece 0 = -50/level, piece 9 = -5/level, piece 10+ = 0. Teaches model to keep board low early, relaxes later. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>9dd7378 -
feat: add one-time height breach penalty (-50 per level above 4)
First time the board height exceeds each level above 4, a -50 penalty is applied. This teaches the model to keep the board low. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>5a39470 -
fix: reduce step penalty from -1 to -0.1
Step penalty of -1 was punishing the model for surviving longer, making death the optimal strategy. Reduced to -0.1. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>2ced530 -
fix: remove height penalty, hole penalty only for new holes
Height penalty removed entirely. Hole penalty now only applies when a step creates new holes, not accumulated every step. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>49913c2
-
Upload src/tetris_env/server/game_engine.py with huggingface_hub
Verified signaturefe115bf