Bug fix for A2C

This commit is contained in:
Antonin Raffin 2020-03-10 17:08:39 +01:00
parent 20ee8cb68d
commit f159a4a9f2

View file

@ -97,8 +97,9 @@ class A2C(PPO):
values = values.flatten()
# Normalize advantage (not present in the original implementation)
advantages = rollout_data.advantages
if self.normalize_advantage:
advantages = (rollout_data.advantages - rollout_data.advantages.mean()) / (rollout_data.advantages.std() + 1e-8)
advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-8)
# Policy gradient loss
policy_loss = -(advantages * log_prob).mean()