DRL-implementations

Implementation of DRL algorithms for solving gym environments. Setup dependencies by installing spinningup. There are baseline implementations of some DRL algorithms in spinningup.

run examples

# can use command line to modify the parameters of dqn agent
# use tensorboard to visualize the results
# python example/dqn.py --name <configuration-name> --render 
python <name-of-the-algorithm>.py --name <configuration-name> --render 

run trained agent

python run_drl.py --name <configuration-name> --render --agent <name-of-agent> --ckpt <ckpt-filepath-of-agent>

remarks

• plotting_utils.py modified from https://github.com/NVIDIA/tacotron2/blob/dd49ffa85085d33f5c9c6cb6dc60ad78aa00fc04/logger.py
• Trainer and Agent only works for discrete action space
• Do not support pixel observation yet
• nec_agent is still testing

TODO

1. unittests