Closed toshima1051 closed 1 month ago
The episode/score
summary shows the return and it seems to be learning well!
The epstats/reward_rate
summary is a measure of how dense the reward is and unrelated.
Thank you very much! I'm glad you are learning successfully, I had no idea that epstats/reward_rate was an irrelevant value! I will use this success as a springboard for further machine learning with Dreanerv3 in the future. I am a first time learner and may have more questions, but I would appreciate your help at that time! Thank you very much!
Hi DreamerV3 team,
I'm currently using DreamerV3 to train an agent on the CartPole-v1 environment using the Gymnasium interface. However, I've encountered an issue where the rewards do not seem to be correctly processed, resulting in poor training performance. Here are the details:
Environment:
Code:
train_dreamerv3.py
:from_gym.py:
Issue: Despite following the standard implementation, the reward seems to be consistently zero, leading to poor learning performance. Below are the TensorBoard logs showing the metrics during training:
Steps to Reproduce:
Execute train_dreamerv3.py with the provided code. Observe the reward metrics in TensorBoard. Expected Behavior: The agent should receive and optimize based on the reward values provided by the environment, resulting in improved episode lengths and scores over time.
Actual Behavior: The rewards remain zero, and the agent does not seem to learn effectively.
Any guidance or suggestions on resolving this issue would be greatly appreciated.
Thank you!