hiro/ray
1
0
Fork 0
mirror of https://github.com/vale981/ray synced 2025-03-05 10:01:43 -05:00
Code Issues Projects Releases Packages Wiki Activity Actions

[RLlib] Fixes CRR flakeyness (#26770)

Browse source
This commit is contained in:
kourosh hakhamaneshi 2022-07-20 12:08:57 -07:00 • committed by GitHub
parent 5433c11650
commit aec79afda1
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
3 changed files with 3 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
rllib/algorithms/crr/torch/crr_torch_policy.py
View file

@ -248,7 +248,7 @@ class CRRTorchPolicy(TorchPolicyV2, TargetNetworkMixin):
q_vals = torch.minimum(q_vals, q_twins)
probs = pi_s_t.dist.probs
v_t = (q_t * probs).sum(-1, keepdims=True)
v_t = (q_vals * probs).sum(-1, keepdims=True)
else:
policy_actions = pi_s_t.dist.sample((n_action_sample,)) # samples

2
rllib/tuned_examples/crr/cartpole-v0-crr.yaml
View file

@ -36,7 +36,7 @@ cartpole_crr:
evaluation_parallel_to_training: True
# specific to CRR
temperature: 1.0
weight_type: bin
weight_type: exp
advantage_type: mean
max_weight: 20.0
n_action_sample: 4

2
rllib/tuned_examples/crr/cartpole-v0-crr_expectation.yaml
View file

@ -36,7 +36,7 @@ cartpole_crr:
evaluation_parallel_to_training: True
# specific to CRR
temperature: 1.0
weight_type: bin
weight_type: exp
advantage_type: expectation
max_weight: 20.0
n_action_sample: 4