matrixBT/Deep-reinforcement-Learning

Deep Reinforcement Learning for Atari with Improvised Architecture

Deep-reinforcement-Learning

Deep Reinforcement Learning for Atari with Improvised* Architecture (More details will be updated soon)

Requirements:

1. Keras-RL

pip install keras-rl

source: https://github.com/matthiasplappert/keras-rl

2. Tensorflow-gpu

conda create --name tensorflow python=3.5

activate tensorflow
pip install tensorflow-gpu
conda install jupyter
conda install scipy

Check: https://www.tensorflow.org/versions/r0.12/get_started/os_setup

3. OpenAi gym

pip install gym

4. Install atari

pip install "gym[atari]"

• For Windows installation: Check:https://github.com/j8lp/atari-py

5. pyglet 1.2.4

pip install https://pypi.python.org/packages/68/c3/300c6f92b21886b0fe42c13f3a39a06c6cb90c9fbb1b71da85fe59091a7d/pyglet-1.2.4-py3-none-any.whl#md5=08e6404a678f91b4eee85eb33b028d88

Results

Model	Parameters	Depth	Ave Reward	Avg Steps Played	Avg training time
DeepMind	16,88,745	5	184.0	4423	207 sec/Epoch
Improvised	17,83,145	10	343.6	7972	258 sec/Epoch

Papers Followed:

1. Playing Atari with Deep Reinforcement Learning
2. Human-level Control Through Deep Reinforcement Learning
3. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size
4. End-to-End Learning to Steer for Self-Driving Cars with Small and Efficient Networks

Trained on: Intel® Xeon® Processor E5, 2.40 GHz, Nvidia Quadro K4200

Author: Bhartendu, Machine Learning & Computing

Mathworks File-exchange LinkedIn