Playing Atari Breakout with Deep Reinforcement Learning
HTML-код
- Опубликовано: 10 сен 2024
- A tutorial on how to make an AI / reinforcement learning agent beating human-level performance in Atari Breakout with Keras and Google Colab (Pro)
Original Paper by Mnih et al. - arxiv.org/abs/...
The Code from the video - github.com/drV...
Would love to see an implementation outside of colab for those of us with local gpu resources :)
Great video though!
Thanks for sharing!
I was struggling on how to install the correct dependencies from openai gym documentation. How exactly would one know which commands to use for a game other than breakout for example?
Can you give the weights of the model at the end of the training? Facing troubles training it, google collab is crashing
Send me an e-mail and I'll send you the trained model
General advice is to reduce the buffer size (150-190 k) and to upgrade to Colab PRO:)
It’s a great example RL. What do you think about Knapsack problem? Can RL solve this problem?
Thank you! According to Wikipedia, you can do it with dynamic programing :) so it's very close to RL
Are there any methods applying policy gradients to solve this problem?
Great Video! Two questions:
Can you somehow send me the trained model?
And how did you get the video, because the only thing I get, is a json containing an empty array under the "video" property.
sure, PM me your mail - I'll send the model
if you do exactly what i show - you gonna be able to get the mp4 file
Or do you have twitter? I cant find a way to contact you on youtube. Maybe you forgot to add it to the "about" section.
@@YouWatchGames I've uploaded the learned model to github!
Thank u for the great tutorial! May I ask which Python version did u use in this one?
3.7.15 (the default one with colab)
Thanks for the video. Am wondering, can this be applied to the game skiing?
I actually am not familiar with this game. It depends upon what kind of reward signal and actions are applied.
paperswithcode.com/sota/atari-games-on-atari-2600-skiing
you can look here - it looks like Q-learning is not in the list of solutions