This repository is a comprehensive implementation of the state-of-the-art Deep Distributional Reinforcement Learning Algorithms that have been introduced as a series of improvements upon each other and are mainly focused on the Atari benchmark.

Included algorithms via a single API:

• C51 (Categorical DQN)
• IQN (Implicit Quantile Networks)
• QRDQN (Quantile Regression DQN)
• FQF (Fully parameterized Quantile Function)

IQN Agent:

Environment: AssaultNoFrameskip-v4

• As it's been highlighted in the part 4 of the FQF paper, FQF algorithm is roughly 20% slower (in terms of runtime speed) than other methods thus, to have a comparable baseline of each method's performance against others, I have plotted the performance metrics with respect to the processing time rather than steps taken by the agent in the environment.
• The runtime was bounded to 6 hours available for free-gpu-equipped machines in the paperspace.com as the host of running the current code.

• gym == 0.19.0
• numpy == 1.22.0
• opencv_python == 4.5.5.62
• psutil == 5.8.0
• torch == 1.13
• wandb == 0.12.9
• protobuf == 3.20
• gym[atari]

python main.py --agent_name="C51" --online_wandb --interval=100 --env_name="BreakoutNoFrameskip-v4"

usage: Choose your desired parameters [-h] [--agent_name AGENT_NAME]
                                      [--env_name ENV_NAME]
                                      [--mem_size MEM_SIZE] [--seed SEED]
                                      [--interval INTERVAL] [--do_test]
                                      [--online_wandb]

optional arguments:
  -h, --help            show this help message and exit
  --agent_name AGENT_NAME
                        Distributional method name
  --env_name ENV_NAME   Name of the environment.
  --mem_size MEM_SIZE   The memory size.
  --seed SEED           The random seed.
  --interval INTERVAL   The interval specifies how often different parameters
                        should be saved and printed, counted by the number of
                        episodes.
  --do_test             The flag determines whether to train the agent or play
                        with it.
  --online_wandb        Run wandb in online mode.

• Training takes significant time. Preferred to be run on machines with access to GPUs.
• Access to wandb allows your run info (as well as graphs) to be saved online.
• Use nohup (or something similar) to execute it in the background.

• Accepted values for agent_name: {"C51", "IQN", "FQF" and "QRDQN"}.
• At the time of testing, the code by default uses the weights of the latest run available in weights folder so, please bear in mind to put your desired weights in the appropriate folder inside the weights directory! 👇

common/logger.py:

def load_weights(self):
    model_dir = glob.glob("weights/*")
    model_dir.sort()
    # model_dir[-1] -> means latest run!
    checkpoint = torch.load(model_dir[-1] + "/params.pth")
    self.log_dir = model_dir[-1].split(os.sep)[-1]
    return checkpoint

1. A Distributional Perspective on Reinforcement Learning, Bellemare, et al., 2017
2. Implicit Quantile Networks for Distributional Reinforcement Learning, Dabney et al., 2018
3. Distributional Reinforcement Learning with Quantile Regression, Dabney et al., 2017
4. Fully Parameterized Quantile Function for Distributional Reinforcement Learning, Yang et al., 2019
5. Distributional Reinforcement Learning (draft), Bellemare et al., 2022

Following repositories were great guides to implement distributional rl ideas. Big thanks to them for their works:

1. DeepRL_PyTorch by @Kchu
2. FQF by @microsoft
3. fqf-iqn-qrdqn.pytorch by @ku2482