code #13
code #13
Conversation
There was a problem hiding this comment.
Interesting.
First, could you please change the name to 09_?_nes_env_name.py
This is a local/random search algorithm, so it should be in 09s. ('10_xx' is for actor-critic methods)
However, I'm not so sure about using the low-level API for the Pong.
It could have been done in a single line with the high-level API. If @hunkim is okay, then I guess it's okay.
I left some comments only in the bipedal example, but the relevant issues in the Pong example should be fixed as well.
|
|
||
| import gym | ||
| import numpy as np | ||
| import cPickle as pickle |
There was a problem hiding this comment.
Are you using Python2?
cPickle is renamed to pickle in python3
no need to import cPickle
| env = gym.make('BipedalWalker-v2') | ||
| np.random.seed(10) | ||
|
|
||
| hl_size = 100 |
There was a problem hiding this comment.
Can you add comments for each hyper parameters?
Also please follow some standard naming conventions.
aver_reward should be renamed to something like avg_reward
| model[k] = model[k] + alpha/(npop*sigma) * np.dot(N[k].transpose(1, 2, 0), A) | ||
|
|
||
| cur_reward = f(model) | ||
| aver_reward = aver_reward * 0.9 + cur_reward * 0.1 if aver_reward != None else cur_reward |
There was a problem hiding this comment.
Why do we need aver_reward?
Why did you use EMA?
|
"high-level API" preferred. |
Added two evolution strategy implementations