Skip to content

Commit 388c8ed

Browse files
committed
Use spike decay for unet
Even 1e-4 spiked the KL
1 parent 6a93092 commit 388c8ed

File tree

1 file changed

+4
-2
lines changed

1 file changed

+4
-2
lines changed

rl_algo_impls/hyperparams/ppo.yml

Lines changed: 4 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -314,7 +314,8 @@ unet-MicrortsDefeatRandomEnemySparseReward-v3:
314314
v_hidden_sizes: [256, 128]
315315
algo_hyperparams:
316316
<<: *microrts-ai-algo-defaults
317-
learning_rate: !!float 1e-4
317+
learning_rate: !!float 2.5e-4
318+
learning_rate_decay: spike
318319

319320
MicrortsDefeatCoacAIShaped-v3: &microrts-coacai-defaults
320321
<<: *microrts-ai-defaults
@@ -356,7 +357,8 @@ unet-MicrortsDefeatCoacAIShaped-v3-diverseBots:
356357
v_hidden_sizes: [256, 128]
357358
algo_hyperparams:
358359
<<: *microrts-ai-algo-defaults
359-
learning_rate: !!float 1e-4
360+
learning_rate: !!float 2.5e-4
361+
learning_rate_decay: spike
360362

361363
HalfCheetahBulletEnv-v0: &pybullet-defaults
362364
n_timesteps: !!float 2e6

0 commit comments

Comments
 (0)