PyTorch RL Minimal Implementations

There are implementations of some reinforcement learning algorithms, whose characteristics are as follow:

  1. Less packages-based: Only PyTorch and Gym, for building neural networks and testing algorithms’ performance respectively, are necessary to install.
  2. Independent implementation: All RL algorithms are implemented in separate files, which facilitates to understand their processes and modify them to adapt to other tasks.
  3. Various expansion configurations: It’s convenient to configure various parameters and tools, such as reward normalization, advantage normalization, tensorboard, tqdm and so on.

RL Algorithms List

Name Type Estimator Paper File
Q-Learning Value-based / Off policy TD Watkins et al. Q-Learning. Machine Learning, 1992
REINFORCE Policy-based On policy MC Sutton et al. Policy Gradient Methods for Reinforcement Learning with Function Approximation. In NeurIPS, 2000.
DQN Value-based / Off policy TD Mnih et al. Human-level control through deep reinforcement learning. Nature, 2015. doing
A2C Actor-Critic / On policy n-step TD Mnih et al. Asynchronous Methods for Deep Reinforcement Learning. In ICML, 2016.
A3C Actor-Critic / On policy n-step TD .Mnih et al. Asynchronous Methods for Deep Reinforcement Learning. In ICML, 2016
ACER Actor-Critic / On policy GAE Wang et al. Sample Efficient Actor-Critic with Experience Replay. In ICLR, 2017. doing
ACKTR Actor-Critic / On policy GAE Wu et al. Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation. In NeurIPS, 2017. doing
PPO Actor-Critic / On policy GAE Schulman et al. Proximal Policy Optimization Algorithms. arXiv, 2017.

Quick Start



tensorboard  # for summary writer
tqdm         # for process bar

Abstract Agent

Components / Parameters

Component Description
policy neural network model
gamma discount factor of cumulative reward
lr learning rate. i.e. lr_actor, lr_critic
lr_decay weight decay to schedule the learning rate
lr_scheduler scheduler for the learning rate
coef_critic_loss coefficient of critic loss
coef_entropy_loss coefficient of entropy loss
writer summary writer to record information
buffer replay buffer to store historical trajectories
use_cuda use GPU
clip_grad gradients clipping
max_grad_norm maximum norm of gradients clipped
norm_advantage advantage normalization
open_tb open summary writer
open_tqdm open process bar


Methods Description
preprocess_obs() preprocess observation before input into the neural network
select_action() use actor network to select an action based on the policy distribution.
estimate_obs() use critic network to estimate the value of observation
update() update the parameter by calculate losses and gradients
train() set the neural network to train mode
eval() set the neural network to evaluate mode
save() save the model parameters
load() load the model parameters

Update & To-do & Limitations

Update History

  • 2021-12-09 ADD TRICK:norm_critic_loss in PPO
  • 2021-12-09 ADD PARAM: coef_critic_loss, coef_entropy_loss, log_step
  • 2021-12-07 ADD ALGO: A3C
  • 2021-12-05 ADD ALGO: PPO
  • 2021-11-28 ADD ALGO: A2C
  • 2021-11-20 ADD ALGO: Q learning, Reinforce

To-do List

  • ADD ALGO DQN, Double DQN, Dueling DQN, DDPG
  • ADD NN RNN Mode

Current Limitations

  • Unsupport Vectorized environments
  • Unsupport Continuous action space
  • Unsupport RNN-based model
  • Unsupport Imatation learning

Reference & Acknowledgements


View Github