wassname/ray
1
0
Fork 0
mirror of https://github.com/wassname/ray.git synced 2026-07-28 11:25:04 +08:00
Code Issues Packages Projects Releases Wiki Activity

[RLLib] DDPG (#1685)

Browse Source
This commit is contained in:
alvkao58
2018-04-11 15:08:39 -07:00
committed by Richard Liaw
parent 74162d1492
commit 15a668dd12
17 changed files with 586 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/ray/rllib/optimizers/local_sync_replay.py
+4 -4
View File
@@ -85,17 +85,17 @@ class LocalSyncReplayOptimizer(PolicyOptimizer):
self.train_batch_size)
weights = np.ones_like(rewards)
batch_indexes = - np.ones_like(rewards)
samples = SampleBatch({
"obs": obses_t, "actions": actions, "rewards": rewards,
"new_obs": obses_tp1, "dones": dones, "weights": weights,
"batch_indexes": batch_indexes})
with self.grad_timer:
td_error = self.local_evaluator.compute_apply(samples)["td_error"]
new_priorities = (
np.abs(td_error) + self.prioritized_replay_eps)
info = self.local_evaluator.compute_apply(samples)
if isinstance(self.replay_buffer, PrioritizedReplayBuffer):
td_error = info["td_error"]
new_priorities = (
np.abs(td_error) + self.prioritized_replay_eps)
self.replay_buffer.update_priorities(
samples["batch_indexes"], new_priorities)
self.grad_timer.push_units_processed(samples.count)