wassname/ray
1
0
Fork 0
mirror of https://github.com/wassname/ray.git synced 2026-07-18 12:40:56 +08:00
Code Issues Packages Projects Releases Wiki Activity

[rllib] Use batch.count in async samples optimizer (#2488)

Browse Source 
Using the actual batch size reduces the risk of mis-accounting. Here, we under-counted samples since in truncate_episodes mode we were doubling the batch size by accident in policy_evaluator.
This commit is contained in:
Eric Liang
2018-07-27 16:44:21 -07:00
committed by GitHub
parent 1e6b130b90
commit 24649726dc
5 changed files with 43 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/ray/rllib/agents/dqn/apex.py
-1
View File
@@ -23,7 +23,6 @@ APEX_DEFAULT_CONFIG = merge_dicts(
"learning_starts": 50000,
"train_batch_size": 512,
"sample_batch_size": 50,
"max_weight_sync_delay": 400,
"target_network_update_freq": 500000,
"timesteps_per_iteration": 25000,
"per_worker_exploration": True,