Physics control tasks with Deep Reinforcement Learning

tutorial

#1

In this tutorial we will implement the paper Continuous Control with Deep Reinforcement Learning, published by Google DeepMind and presented as a conference paper at ICRL 2016. The networks will be implemented in PyTorch and using OpenAI gym. The algorithm combines Deep Learning and Reinforcement Learning techniques to deal with high-dimensional, i.e. continuous, action spaces.


This is a companion discussion topic for the original entry at https://blog.paperspace.com/physics-control-tasks-with-deep-reinforcement-learning

#2

Hi Antonio,
Great job man, it’s really helped me to get through the RL as I’m new to this concept and in general field of AI.
So, I faced some issue when tried to run the code on the CPU. Sadly, I don’t have an external GPU to run the process on. :expressionless:
Anyway, here is the output of the last block of code:

---------------------------------------------------------------------------
NotImplementedError                       Traceback (most recent call last)
<ipython-input-24-97b79fb15b62> in <module>
     31         a += noise()*max(0, epsilon)
     32         a = np.clip(a, -1., 1.)
---> 33         s2, reward, terminal, info = env.step(a)
     34 
     35 

~/anaconda3/envs/lstm/lib/python3.7/site-packages/gym/core.py in step(self, action)
    283 
    284     def step(self, action):
--> 285         return self.env.step(self.action(action))
    286 
    287     def action(self, action):

~/anaconda3/envs/lstm/lib/python3.7/site-packages/gym/core.py in action(self, action)
    286 
    287     def action(self, action):
--> 288         raise NotImplementedError
    289 
    290     def reverse_action(self, action):

NotImplementedError: 

I tried to adjust the code to run on CPU :grin: and succeeded to fix some of it but apparently there should be more to consider. :sweat_smile: Any suggestions highly appreciated. :v: