ByungSunBae

# from : https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/graph_editor/examples/edit_graph_example.py
import numpy as np

import tensorflow as tf
from tensorflow.contrib import graph_editor as ge

# create a graph
g = tf.Graph()

with g.as_default():

ByungSunBae

## Convex Optimization in R using CVXR
## from : https://rviews.rstudio.com/2017/11/27/introduction-to-cvxr/

## load CVXR
if (!require(CVXR)){
  install.packages("CVXR")
} else{
  require(CVXR)
}

ByungSunBae

# 2) RNN model 생성----------------
## Refer : https://github.com/hunkim/DeepLearningZeroToAll/blob/master/lab-12-5-rnn_stock_prediction.py
library(tensorflow)
library(reticulate)

contrib <- tf$contrib

tf$reset_default_graph()

# train Parameters

ByungSunBae

DRL : Policy Gradient

Overview

• Deep Q network(이하 DQN)와 그 Variants(DoubleDQN, DuelingDQN)를 구현해보았다.

---

• 그런데 DQN과 같은 계열의 방법은 RL에서 Value-Based Method라서 Policy에 대한 직접적인 학습이 아니다.

• 게다가 Value가 약간만 바뀌어도 Policy가 금방 변한다.

• 학습과정이 불안정하게되어 수렴자체가 불안정해진다. (Bias가 높다.)

ByungSunBae

# Thanks for 주찬웅
# References:
## 1) https://github.com/jcwleo/Reinforcement_Learning/blob/master/Breakout/
## 2) http://pytorch.org/tutorials/ 
## 3) https://github.com/transedward/pytorch-dqn

## My codes is very very dirty...
## I want to your idea and advice that improves this codes.

import argparse