jesusbaez / lunarlander-v2 Goto Github PK

View Code? Open in Web Editor NEW

This project aims to learn an agent to guide a space vehicle from starting point to the landing pad without crashing in OpenAI's gym environment called LunarLander-v2

License: GNU General Public License v3.0

HTML 95.91% TeX 1.84% Python 2.25%

lunarlander-v2's Introduction

Capstone project - LunarLander-v2

Problem statement

LunarLander-v2 is an environment in OpenAI's gym package. The aim is to learn an agent to guide a space vehicle from starting point to the landing pad without crashing. The state space is R⁸ and there are four possible actions { do nothing, fire left orientation engine, fire main engine, fire right orientation engine }.

Software requirements

python 2.7
numpy
OpenAI's gym package
Keras with tensorflow or theano

Soft dependency - no need to install in the standard way

Keras-rl package - required for running the high benchmark agent.

Setting up the project :

$ git clone https://github.com/matthiasplappert/keras-rl.git

$ git clone https://github.com/dennisfrancis/LunarLander-v2.git

$ cd keras-rl

$ git am < ../LunarLander-v2/keras-rl-testrun-fix.patch   # Apply a patch for keras-rl to work with the project

$ cd ../LunarLander-v2/src

$ ln -s ../../keras-rl/rl rl   # create a symlink to rl dir of keras-rl to src dir

Running the agents

$ cd src                          # Go to source code directory

$ rm -rf ../monitor               # Required only if you have run some of the agents before

$ python run_random_agent.py      # To run random agent or low benchmark

$ python run_high_benchmark.py    # To run high benchmark agent

$ python run_basic_dqn.py         # To run basic DQN agent

$ python run_full_dqn.py          # To run improved DQN agent