search

bilalkabas / DRL-Nav

Deep Reinforcement Learning based autonomous navigation in realistic simulation environments.
GNU Affero General Public License v3.0
32 stars 5 forks source link

about episode and the step size of each episode defined、plot #2

Closed xiaomaohouhou closed 2 years ago

xiaomaohouhou commented 2 years ago

Hi, I'd like to consult you two questions. Where are the number of episode and the step size of each episode defined? Can you share the plot file of the result(such as loss、mean reward、success rate )? Thank you very much, my friend

bilalkabas commented 2 years ago

Hi,

  1. The step size of an episode varies depending on the agent's actions, i.e. the episode is terminated when the drone collides with walls or it goes through a hole successfully. Therefore, there is no fixed step size for episodes. Training is performed over a specific number of steps not episodes. In below lines, the total_timesteps refer to the number of steps to be reached to finish training.

https://github.com/bilalkabas/DRL-Nav/blob/11089e02141c6373eb5fe6b70c107ce27090d5b0/train.py#L75-L79

  1. For the mean reward per episode, see the below figure.

The loss is not so descriptive for reinforcement learning and I cannot find it at the moment. For the success rate, I do not have a direct graph but it would look like the mean episode reward as the reward computation includes success and fail scenarios.