hiro/ray
1
0
Fork 0
mirror of https://github.com/vale981/ray synced 2025-03-06 02:21:39 -05:00
Code Issues Projects Releases Packages Wiki Activity Actions

[RLlib] Fix combination of lockstep and multiple agnts controlled by the same policy. (#9521)

Browse source 
* Change aggregation when lockstep is activated.

Modification of MultiAgentBatch.timeslices to support the combination of lockstep and multiple agents controlled by the same policy.

fix ray-project/ray#9295

* Line too long.
This commit is contained in:
Raphael Avalos 2020-07-20 08:03:12 +02:00 • committed by GitHub
parent 0cee75c86a
commit 440c9c42be
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 3 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
rllib/policy/sample_batch.py
View file

@ -438,8 +438,8 @@ class MultiAgentBatch:
steps = []
for policy_id, batch in self.policy_batches.items():
for row in batch.rows():
steps.append((row[SampleBatch.EPS_ID], row["t"], policy_id,
row))
steps.append((row[SampleBatch.EPS_ID], row["t"],
row["agent_index"], policy_id, row))
steps.sort()
finished_slices = []
@ -458,7 +458,7 @@ class MultiAgentBatch:
# For each unique env timestep.
for _, group in itertools.groupby(steps, lambda x: x[:2]):
# Accumulate into the current slice.
for _, _, policy_id, row in group:
for _, _, _, policy_id, row in group:
cur_slice[policy_id].add_values(**row)
cur_slice_size += 1
# Slice has reached target number of env steps.