Curiosity Driven Deep Reinforcement Learning
Deep Reinforcement Learning
0% Not started
Paper Analysis: Asynchronous Methods for Deep Reinforcement Learning
Paper Analysis: Curiosity Driven Exploration Self Supervised Prediction
In 2012 I received my PhD in experimental condensed matter physics from West Virginia University. Following that I was a dry etch process engineer for Intel Corporation, where I leveraged big data to make essential process improvements for mission critical products. After leaving Intel in 2015, I have worked as a contract and freelance deep learning and artificial intelligence engineer.
I am trying to work on a similar implementation of this and I really need to set the seeds for reproducibility. I have tried everything but it does not work as expected. I have tried setting seeds for torch, random , numpy , the environment and the environment action space both in the main and worker function but no matter what i cannot achieve reproducibility. Would you have any insight on how I could achieve it ?
The easiest way is to drop a link to your code on github and I can take a look.
Just as a general strategy, what I try to do in these situations is strip out complexity. Start with a very basic script where you just generate random numbers / reset the environment. Make sure you can get reproducibality there, and then compare to the more complex code.
Thank you for the reply.
Yes I might try doing in a more simple implementation, but the problem i think lies in the shared memory remember function or when loading the model with state_dict .
I tried fixing the seeds on your simple cartpole implementation without the encoders but I couldnt even do it there.
This is my repository https://github.com/katiavas/rep.git