OpenAI-Gym-NeuralNetBot

Deep Neural Network that learns to play OpenAI gym's games Uses Tensorflow library for the DNN model.

• Starts by playing the game 100000 times (gamesToTrain variable) only with possible random moves.
• Adds the game data ([observation, action]) from games that finished with a score over 100 (scoreMinimum) to the training data.
• Trains with that data.
• Plays the game 100 times (modelPlayNumber) with the DNN's prediction (with the current observation as input).
• Adds the new game data to the training data.
• Saves it so the next time it runs it can learn from more and better data.

Requirements

• Gym
• Tensorflow / Tensorflow-gpu
  • CUDA Toolkit 8.0 (optional - to use with the gpu version)
  • cuDNN 7 (optional - to use with the gpu version)
• TFlearn
• Numpy

How To Use

#Set ENV_NAME in OpenAI_Bot_CartPoleExample.py:15 to the name of environment to learn
ENV_NAME = 'CartPole-v1'  

#Uncomment OpenAI_Bot_CartPoleExample.py:118 if you want to render the game environment 
#as the model is playing it (after being trained)
env.render() #Renders the environment to see what is happening.  

#Comment OpenAI_Bot_CartPoleExample.py:172 if you don't want to test if the
#environment is the correct one and if it's working properly
playGamesWithRandomActions()  

#Uncomment OpenAI_Bot_CartPoleExample.py:173 when you wish to train the neural
#network and run it
modelTrainAndPlay()

#Run OpenAI_Bot_CartPoleExample.py to train/load the model.
$ python OpenAI_Bot_CartPoleExample.py