Skip to content

Commit f5b98ca

Browse files
author
Ervin T
authored
Improved SAC hyperparameters for Crawler, Walker (#2635)
* Tweak SAC hyperparams * Make network bigger * Properly report entropy * Revert "Properly report entropy" This reverts commit 383a8d8.
1 parent 62e8fb1 commit f5b98ca

File tree

1 file changed

+12
-4
lines changed

1 file changed

+12
-4
lines changed

config/sac_trainer_config.yaml

Lines changed: 12 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -159,25 +159,33 @@ CrawlerStaticLearning:
159159
normalize: true
160160
time_horizon: 1000
161161
batch_size: 256
162-
train_interval: 3
162+
train_interval: 2
163163
buffer_size: 500000
164164
buffer_init_steps: 2000
165165
max_steps: 5e5
166166
summary_freq: 3000
167167
init_entcoef: 1.0
168168
num_layers: 3
169169
hidden_units: 512
170+
reward_signals:
171+
extrinsic:
172+
strength: 1.0
173+
gamma: 0.995
170174

171175
CrawlerDynamicLearning:
172176
normalize: true
173177
time_horizon: 1000
174178
batch_size: 256
175179
buffer_size: 500000
176180
summary_freq: 3000
177-
train_interval: 3
181+
train_interval: 2
178182
num_layers: 3
179183
max_steps: 1e6
180184
hidden_units: 512
185+
reward_signals:
186+
extrinsic:
187+
strength: 1.0
188+
gamma: 0.995
181189

182190
WalkerLearning:
183191
normalize: true
@@ -186,8 +194,8 @@ WalkerLearning:
186194
buffer_size: 500000
187195
max_steps: 2e6
188196
summary_freq: 3000
189-
num_layers: 3
190-
train_interval: 3
197+
num_layers: 4
198+
train_interval: 2
191199
hidden_units: 512
192200
reward_signals:
193201
extrinsic:

0 commit comments

Comments
 (0)