SirStone
fedab54bc0
- Accumulate transitions across 10 rounds (~3000) before PPO update (was per-round ~300 — gradient estimates were far too noisy) - training.nim: MAX_TRANSITIONS 4096→8192, done flag on transitions, GAE handles episode boundaries correctly - PPO_Bot.nim: buffer persists across rounds, update every N rounds - training.env: lr 5e-5→1e-4, entropy 0.001, UPDATE_INTERVAL=10
19 lines
478 B
Bash
19 lines
478 B
Bash
# TRAINING_OPPONENT=Fire
|
|
TRAINING_ROUNDS=30000
|
|
PPOB_LOG_FILE=/home/davide/Projects/SirRoboGarage/tools/training_runner/logs/fire_training.jsonl
|
|
PPOB_LR=1e-4
|
|
PPOB_UPDATE_INTERVAL=10
|
|
PPOB_CLIP_EPSILON=0.2
|
|
PPOB_ENTROPY_COEFF=0.001
|
|
PPOB_VALUE_LOSS_COEFF=0.5
|
|
PPOB_MAX_GRAD_NORM=0.5
|
|
PPOB_GAMMA=0.99
|
|
PPOB_LAM=0.95
|
|
PPOB_EPOCHS=4
|
|
PPOB_MINI_BATCH_SIZE=64
|
|
PPOB_LOG_STD_FLOOR=-3.0
|
|
PPOB_LOG_STD_CEILING=0.0
|
|
PPOB_INITIAL_LOG_STD=-0.5
|
|
|
|
# PPOB_EVAL_ONLY=1 → freeze training (pure evaluation)
|