BipedalWalker
Branching DQN implemetation with pytorch in BipedalWalker environment. No PER, No DDQN version to make it easier for you to understand the point. And It is also sufficiently capable of showing (almost) optimal movements after 1000 epochs.
RUN
python main.py
-
if you want to change hyper-parameters, you can check "python main.py --help"
-
you just train and test basic model using main.py
-
'--train', type=bool, default=True, help="(default: True)"
-
'--render', type=bool, default=False, help="(default: False)"
-
'--epochs', type=int, default=1000, help='number of epochs, (default: 1000)'
-
'--tensorboard', type=bool, default=False, help='use_tensorboard, (default: False)'
-
'--lr_rate', type=float, default=0.0001, help='learning rate (default : 0.0001)'
-
'--batch_size', type=int, default=64, help='batch size(default : 64)'
-
'--gamma', type=float, default=0.99, help='gamma (default : 0.99)'
-
'--action_scale', type=int, default=6, help='action scale between -1 ~ +1'
-
"--load", type=str, default = 'no', help = 'load network name in ./model_weights'
-
"--save_interval", type=int, default = 100, help = 'save interval(default: 100)'
-
"--print_interval", type=int, default = 1, help = 'print interval(default : 1)'
Install BipedalWalker
Ubuntu
conda install swig # needed to build Box2D in the pip install
pip install box2d-py # a repackaged version of pybox2d
https://stackoverflow.com/questions/44198228/install-pybox2d-for-python-3-6-with-conda-4-3-21