Nusha97/Deep-RL-for-Atari-Games

Investigating different reinforcement algorithms on openai baseline datasets

Deep-RL-for-Atari-Games

Investigating different reinforcement algorithms on openai baseline datasets

We actually use two platforms according to our accessible resources:

Notebooks is used for training CartPole with DQN and both games with A2C (run on colab)

baselines is used for training Acrobot with DQN (run on server with GPU)

Note for baselines

results (including both results of DQN and A2C) stored in ./logs folder; pics stored in ./pics folder

Changed files:

setup.py: cloudpickle==1.2.0 //I have this version problem. If you don't have then it can be ignored.

train_xxx.py: logger.configure("logs/xxx") //to set the result saving path

run.sh //to run the training

plot.py //to plot results

To run:

setup environment as baselines readme file (setup.py has to be changed before running pip install -e . to install right version)

tuning parameters in ./baselines/deepq/defaults.py or build a file like ./baselines/deepq/experiments/train_cartpole.py

sh ./run.sh

python plot.py

Games to test:

CartPole-v0, Acrobot-v1

Parameters to be tuned:

learning rate, batch size, buffer_size