README.md

Environment PongNoFrameskip-v4

Goals of the repo:

• Demystify RL algorithms by providing minimal, pytorch object-oriented implementations and it's accompanying pseudocode and explanation
  • I also provide quick explanations on typical Pytorch tricky manipulations, like .squeeze() or .detach() (at the end of repo)
• Support my theory notes
• Practice implementing algorithms

Features:

• Accompanying theory notes 📕
• Minimal and Object Oriented code for simple (Semi Gradient Sarsa) and state of the art algorithms (PPO-Clip)
• Understandable and intuitive logging via experience tracking
• Easy reward and loss plotting
• Hyperparameter Tuning (in < 40 lines of code for all algorithms)*
• Intuitive terminal interface (in < 50 lines)

Disclaimer:

• These implementations aren't supposed to be used in research, but for full-transparency learning.
  As so, no testing capabilities or pre-trained models are provided.

Algorithms

Algorithm	Lines of Code	Verified Environments
Semi Gradient Sarsa	< 100	CartPole-v1
Reinforce with Baseline	< 100	CartPole-v1
Deep Deterministic Policy Gradient (DDPG)	~ 150	HalfCheetah-v2 , Pendulum-v1
N Step Actor Critic	< 150	CartPole-v1, LunarLander-v2
Double Deep Q Network (DDQN)	< 200	CartPole-v1, LunarLander-v2, PongNoFrameskip-v4
Proximal Policy Optimization (PPO)	< 200	CartPole-v1, LunarLander-v2, Pendulum-v1

Usage

For a complete description run

pyhton run.py -h

Train

python run.py --algo <algo> --env <env>

Outputs

<algorithm>
  └─logs
    └─<environment> 
      └─<experiment_id>
          ├─config.txt >> Contains agent configuration 
          ├─log.txt >> Stdout output (useful for customization)
          └─results.csv >> CSV of Rewards and Loss

For example running

python run.py --algo ddqn --env CartPole-v1

will yield

ddqn
  └─logs
    └─CartPole-v1 
      └─experiment_0
          ├─config.txt >> Contains agent configuration 
          ├─log.txt >> Stdout output (useful for customization)
          └─results.csv >> CSV of Rewards and Loss

• Running again does not overwrite, but appends new experiments <algo>/logs/<env> folder
• The files begin to be written as soon as the experiment starts. Hence interrupting via CTRL+C will still yield plottable results.
• You can also delete last experiment by running the same command with -d or --delete flag

Tune / Optimize

This is mostly helpful if you plan on adding different environments. It uses optuna to run several hyperparameter combinations and picks the best.

python run.py --algo ddpg --env Pendulum-v1 --optimize --n-trials 100

Outputs

<algorithm>
  └─logs
    └─<environment>
      └─<experiment>
        └─trial_1
          ├─config.txt >> Contains agent configuration 
          ├─log.txt >> Stdout output (useful for customization)
          └─results.csv >> CSV of Rewards and Loss
        └─trial_2
          └─...
        └─...
        └─trial_<n_trials>
          └─...

• You can cancel it with CTRL+C
• For simplicity, all hyperparameter suggestions are done in core.optuna_create method. I'll leave the tweaking around for you.

Plot

Used for plotting losses and rewards

python run.py --algo <algo> --env <env> --plot <experiment>

Ex: python run.py --algo ppo --env CarPole-v1 --plot experiment_6

Would open


• <experiment> can be ommited and it will use latest experience for specified algorithm and environment.

TODO

• put pseudocde image into every folder
• run every algorithm and put some graphs
• Add at least 2 environments per main algorithm
• unify all train.py
• consistent signature across all algorithms
• Add Atari onto DQN
• add logging
  • To csv
• explain folder structure
• experiment manager
• add no test disclaimer
• separate plotting from main execution (use intermidiary csv)
• vectorized environments. (Check stable_baselines3/common/vec_env)
• unit tests to verify all algos work (take a look at rl-baselines3-zoo/tests)
• Allow for video saving
• Allow for model saving

other improvements

• Create Environment class with run_episode code

troubleshooting

• In case you run into ROM license troubles when running PongNoFrameskip-v4, run

pip install "gym[atari,accept-rom-license]"

Be aware that this accepts the ROM license for you.

Pytorch tricks

• .squeeze(...)
  • This "squeezes" dim=1 of the array. Useful for when you're working with slightly different tensor shapes. Ex: tensor of dims [32,1].squeeze() -> [32,]
• .unsqueeze(...)
  • Kind of the reverse of .squeeze(), as it adds in one dimension. Ex: tensor of dims [32,].unsqueeze(1) -> [32, 1]
• .view(...)
  • Like .unsqueeze() but you could do weirder manipulations (change several dimensions). Ex tensor of shape [2,2] .view(4,1) -> [4,1]
• detach()
  • Detaches tensor from gradient calculations. Useful when you want to make predicitons, without backpropagating. ( Ex: DDQN target network)