parameters values about atari-skiing-rl HOT 4 CLOSED

from atari-skiing-rl.

Hi,
Do you mean the gif animation?
After searching for the optimal hyperparameters, I trained the agent using the following arguments:
skiing.py --replay_memory 5000 --agent_history 5 --downsample 2 --frame_skipping 3 --fit_frequency 3 --final_epsilon 0.1 --decay 1e-5 --observe 5e4 --batch 64 --learning_rate 0.95
The --no_render argument was also used in order to train the model fast and not have to wait for the graphics to render.
After the training had finished, I used the trained agent (--agent argument) and ran the script only for one episode, using the argument --record in order to record it and --recording_name_prefix in order to name the file as trained-agent.
Finally, I recorded the script for an untrained agent without random steps (--no_operation set to 0) and named the recorded file as untrained_agent.
In order to generate the comparison gif, I used a video editing program to sync the two gifs duration and render them side by side.

I just noticed that I have not updated the Readme file to contain the recording arguments, though you can use the --help argument which will show the most recent help description.

from atari-skiing-rl.

from atari-skiing-rl.

Hello, the agent was trained for 10,000 episodes in order to get the demonstrated results.

from atari-skiing-rl.