Mountain car RL task

Env from gym and compare different methods on two tasks: Mountain car v0 and Continuous mountain car v0

Env

• Ubuntu
• Tensorflow 1.x (test on 1.12.0)
• Gym (test on 0.15.4)
• Numpy (test on 1.17.2)
• Matplotlib (test on 3.1.1)

# or if you use anaconda, you can directly run: 
$ conda env create -f environment.yml

Usage

Notice if you want to render the video, in beginning of each file:

RENDER = True  # Show GUI

Mountain Car v0

Result: DQN>Sarsa-lambda>Sarsa~QTable

$ cd MountainCar
# Run QTable
$ python QTable_MountainCar.py
# Run Sarsa
$ python Sarsa_MountainCar.py
# Run Sarsa-lambda
$ python Sarsalambda_MountainCar.py
# Run DQN
$ python DQN_MountainCar.py

Continuous Mountain Car v0

Result: A2C>A3C

$ cd MountainCar_Continuous
# Run A2C
$ python A2C_MountainCar.py
# Run A3C
$ python A3C_MountainCar.py

Citation

@misc{RLYixuan2020, 
    author = {Yixuan, Wei},
    title = {Mountain car RL task},
    howpublished = {\url{https://github.com/weiyx16/RL_Toy_Example}},
    year = {2020}
}

Thanks

• https://github.com/MorvanZhou/Reinforcement-learning-with-tensorflow
• https://towardsdatascience.com/reinforcement-learning-w-keras-openai-dqns-1eed3a5338c
• https://github.com/stefanbo92/A3C-Continuous
• https://github.com/sudharsan13296/Hands-On-Reinforcement-Learning-With-Python
• https://medium.com/@asteinbach/actor-critic-using-deep-rl-continuous-mountain-car-in-tensorflow-4c1fb2110f7c
• https://towardsdatascience.com/understanding-actor-critic-methods-931b97b6df3f