korymath

Following the installation and first test for torch-twrl
https://github.com/twitter/torch-twrl

Run code:
cd examples
chmod u+x cartpole-pg.sh
./cartpole-pg.sh

korymath

from sklearn.feature_extraction.text import CountVectorizer
from collections import Counter
import numpy as np

with open('accepted-papers.txt', 'r') as myfile:
	data=myfile.read().replace('\n', '')

lectures = data
bigram_vectorizer = CountVectorizer(ngram_range=(2, 10), stop_words='english')
analyze = bigram_vectorizer.build_analyzer()

korymath

"""
Translated from https://webdocs.cs.ualberta.ca/~sutton/MountainCar/MountainCar1.cp
Algorithm described at https://webdocs.cs.ualberta.ca/~sutton/book/ebook/node89.html
Some minor adjustments to constants were made to make the program work on environments
besides Mountain Car.
"""

import random
import math
import numpy as np

korymath

SONNETS



TO THE ONLY BEGETTER OF
THESE INSUING SONNETS
MR. W. H. ALL HAPPINESS
AND THAT ETERNITY
PROMISED BY
OUR EVER-LIVING POET WISHETH

korymath

hs.hotkey.bind({"cmd", "alt", "ctrl"}, "Left", function()
  local win = hs.window.focusedWindow()
  local f = win:frame()
  local screen = win:screen()
  local max = screen:frame()

  f.x = max.x
  f.y = max.y
  f.w = max.w / 2
  f.h = max.h

korymath

"""
Policy Gradients
1. Sample paths.
2. Process paths (compute advantage, baseline, rewards, etc)
3. Run the paths through the policy (function approximator)
4. Compute gradients/update policy model weights
5. Profit?!?!
How we optimize the policy
--------------------------
L(theta) = sum t=0 to T-1 log policy(action_t | state_t, theta) * A_t

korymath

Using the rllab TRPO algorithm

korymath

Using TRPO from https://github.com/rllab/rllab

korymath

Following https://github.com/rllab/rllab

korymath

import gym
env = gym.make('Reacher-v1')
env.reset()
env.render()

env.monitor.start('/tmp/reacher-1', force=True)
for i_episode in xrange(101):
    observation = env.reset()
    for t in xrange(100):
        env.render()