chenglongchen / pytorch-drl

PyTorch implementations of various Deep Reinforcement Learning (DRL) algorithms for both single agent and multi-agent.

pytorch-madrl

This project includes PyTorch implementations of various Deep Reinforcement Learning algorithms for both single agent and multi-agent.

• A2C
• ACKTR
• DQN
• DDPG
• PPO

It is written in a modular way to allow for sharing code between different algorithms. In specific, each algorithm is represented as a learning agent with a unified interface including the following components:

• interact: interact with the environment to collect experience. Taking one step forward and n steps forward are both supported (see _take_one_step_ and _take_n_steps, respectively)
• train: train on a sample batch
• exploration_action: choose an action based on state with random noise added for exploration in training
• action: choose an action based on state for execution
• value: evaluate value for a state-action pair
• evaluation: evaluation the learned agent

Requirements

• gym
• python 3.6
• pytorch

Usage

To train a model:

$ python run_a2c.py

Results

It's extremely difficult to reproduce results for Reinforcement Learning algorithms. Due to different settings, e.g., random seed and hyper parameters etc, you might get different results compared with the followings.

A2C

ACKTR

DDPG

DQN

PPO

TODO

• TRPO
• LOLA
• Parameter noise

Acknowledgments

This project gets inspirations from the following projects:

• Ilya Kostrikov's pytorch-a2c-ppo-acktr (kfac optimizer is taken from here)
• OpenAI's baselines

License

MIT

pytorch-drl's Issues

bug

https://github.com/ChenglongChen/pytorch-madrl/blob/44b8f082ffb054c6dabead8971b66a7be4c1f9ff/A2C.py#L56

should add one more line: self.critic.cuda().

Also could you provide some example for running the multi-agent case? Thanks!

The Actor Critic Structure in MAA2C

A little confused about your implementation of MAA2C. I don't think the input of the actor network is simply the ``joint state" of the agents. According to [1] the critic's input should be state of the environment (where agents' joint state is not necessarily defined) + the joint action of the agents, i.e., the critic here should be a Q-function for joint actions. And for the actor it should be something like a policy, where I am not quite understand why the actor network is implemented in this way. Appreciate if explained.

License?

Hi,

What is the license?

Hugh

how's this framework going in 2021?

About the computation of Advantage and State Value in PPO

In your implementation of Critic, you feed the network of the observation and action and output 1-dim value. Can I make the inference that It is Q(s,a) ?
But the advantage you given is
values = self.critic_target(states_var, actions_var).detach() advantages = rewards_var - values
It is the estimation of q_t minus Q(s_t,a)
I think it should be Advantage = q_t - V(s_t)

No continuous action space mode

cannot find common.kfac

In the ACKTR.pg, a KFACOptimizer is imported. But no source file is found