nazar-ivantsiv

I implemented the DQN model from this paper: https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf.

I used a simple network with two hidden layers and an output layer, instead of the CNN described in the paper due to the relative simplicity of the Cart-Pole environment compared to Atari games.

Note, that I did not yet implement the target network described in the more recent paper here: https://storage.googleapis.com/deepmind-data/assets/papers/DeepMindNature14236Paper.pdf.

The results vary from run to run, sometimes taking 1000 episodes to solve the problem, and at other times taking only 200 episodes.

	# import the gym stuff
	import gym
	# import other stuff
	import random
	import numpy as np
	# import own classes
	from deepq import DeepQ

	env = gym.make('CartPole-v0')

	# import os
	# os.environ["THEANO_FLAGS"] = "mode=FAST_RUN,device=gpu,floatX=float32"
	# import theano

	# import the neural net stuff
	from keras.models import Sequential
	from keras import optimizers
	from keras.layers.core import Dense, Dropout, Activation
	from keras.layers.normalization import BatchNormalization
	from keras.layers.advanced_activations import LeakyReLU

	""" Trains an agent with (stochastic) Policy Gradients on Pong. Uses OpenAI Gym. """
	import numpy as np
	import cPickle as pickle
	import gym

	# hyperparameters
	H = 200 # number of hidden layer neurons
	batch_size = 10 # every how many episodes to do a param update?
	learning_rate = 1e-4
	gamma = 0.99 # discount factor for reward