PyTorch Implementation of Implicit Quantile Networks (IQN) for Distributional Reinforcement Learning with additional extensions like PER, Noisy layer and N-step bootstrapping. Creating a new Rainbow-DQN version. This implementation allows it also to run and train on several environments in parallel!

• Baseline IQN Notebook
• Script Version with all extensions: IQN The IQN Baseline in this repository is already a Double IQN version with target networks!

• Dueling IQN
• Noisy layer
• N-step bootstrapping
• Munchausen RL
• Parallel environments for faster training (wall clock time). For CartPole-v0 3 worker reduced training time to 1/3!

With the script version it is possible to train on simple environments like CartPole-v0 and LunarLander-v2 or on Atari games with image inputs!

To run the script version: python run.py -info iqn_run1

To run the script version on the Atari game Pong: python run.py -env PongNoFrameskip-v4 -info iqn_pong1

To see the options: python run.py -h

-agent, choices=["iqn","iqn+per","noisy_iqn","noisy_iqn+per","dueling","dueling+per", "noisy_dueling","noisy_dueling+per"], Specify which type of IQN agent you want to train, default is IQN - baseline!
-env,  Name of the Environment, default = BreakoutNoFrameskip-v4
-frames, Number of frames to train, default = 10 mio
-eval_every, Evaluate every x frames, default = 250000
-eval_runs, Number of evaluation runs, default = 2
-seed, Random seed to replicate training runs, default = 1
-munchausen, choices=[0,1], Use Munchausen RL loss for training if set to 1 (True), default = 0
-bs, --batch_size, Batch size for updating the DQN, default = 8
-layer_size, Size of the hidden layer, default=512
-n_step, Multistep IQN, default = 1
-N, Number of quantiles, default = 8
-m, --memory_size, Replay memory size, default = 1e5
-lr, Learning rate, default = 2.5e-4
-g, --gamma, Discount factor gamma, default = 0.99
-t, --tau, Soft update parameter tat, default = 1e-3
-eps_frames, Linear annealed frames for Epsilon, default = 1 mio
-min_eps, Final epsilon greedy value, default = 0.01
-info, Name of the training run
-w, --worker, Number of parallel environments. Batch size increases proportional to number of worker. Not recommended to have more than 4 worker, default = 1
-save_model, choices=[0,1]  Specify if the trained network shall be saved or not, default is 0 - not saved!

tensorboard --logdir=runs

Trained and tested on:

Python 3.6 
PyTorch 1.4.0  
Numpy 1.15.2 
gym 0.10.11 

IQN and Extensions (default hyperparameter):

Dueling IQN and Extensions (default hyperparameter):

IQN and M-IQN comparison (only trained for 500000 frames ~ 140 min).

Hyperparameter:

• frames 500000
• eps_frames 75000
• min_eps 0.025
• eval_every 10000
• lr 1e-4
• t 5e-3
• m 15000
• N 32

Performance after 10 mio frames, score 258

• Comparison plot for n-step bootstrapping (n-step bootstrapping with n=3 seems to give a strong boost in learning compared to one step bootstrapping, plots will follow)
• Performance plot for Pong compared with Rainbow
• adding Munchausen RL ☑

Im open for feedback, found bugs, improvements or anything. Just leave me a message or contact me.

• IQN
• Dueling DQN
• Noisy layer
• C51
• PER

• Sebastian Dittert

Feel free to use this code for your own projects or research. For citation:

@misc{IQN and Extensions,
  author = {Dittert, Sebastian},
  title = {Implicit Quantile Networks (IQN) for Distributional Reinforcement Learning and Extensions},
  year = {2020},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/BY571/IQN}},
}