lirnli

## A simple Q-learning net with memory relay
import numpy as np
import tensorflow as tf
from matplotlib import pyplot as plt
import gym

# Get env parameters
GYM_NAME = 'CartPole-v0'
env = gym.make(GYM_NAME)
obs_shape = env.observation_space.shape

lirnli

# Two Q nets are used. They use themsevleves to estimate Q(t,a) but use the other to estimate argmax Q(t+1,a). 
# Still use one-step algorithm in training Q nets.
# Based on arXiv:1509.06461 [cs.LG]
# https://lirnli.wordpress.com/2017/08/17/debugging-reinforcement-neural-network-deep-q-net/
#
# Hyperparameter summary:
#   reward decay rate = 0.999
#   memory relay (with 10000 observations)
#   AdamOptimizer, with learning rate decay
#   tanh activation function

lirnli

DDPG solution
Jupyter notebook file at
https://github.com/lirnli/OpenAI-gym-solutions/blob/master/Continuous_Deep_Deterministic_Policy_Gradient_Net/DDPG%20Class%20ver2.ipynb

lirnli

DDPG solution

Jupyter notebook file at
https://github.com/lirnli/OpenAI-gym-solutions/blob/master/Continuous_Deep_Deterministic_Policy_Gradient_Net/DDPG%20Class%20ver2%20(Pendulum-v0).ipynb

lirnli

lirnli

lirnli

lirnli

lirnli

lirnli