README.md

AGNES - Flexible Reinforcement Learning Framework with PyTorch

Status: This framework is under active development and bugs may occur.

Results

MuJoCo

(Current results) MuJoCo "Ant-v2" training with 1M steps. Single runner with PPO algorithm, MLP NN and 32 number of envs. The curve is an average of 3 runs.

You can get the Tensorboard log file by clicking the image above(You will be redirected to the destination GitHub folder). The default config for the MuJoCo environment was used. Plotted by examples/plot.py

Atari

(Old results)

Atari "BreakoutNoFrameskip-v4" with frame stack training with 10M steps. DistributedMPI runner with PPO algorithm, LSTMCNN and 16 number of envs.

You can get the Tensorboard log file by clicking the image above(You will be redirected to the destination GitHub folder). The default config for the Atari environment was used.

Grad-cam technique was used for sampled action chosen by trained LSTMCNN(previous point).

Runners

Single

One worker and trainer. agnes.make_vec_env can also be used here.

import agnes
import time


if __name__ == '__main__':
    env = agnes.make_env("InvertedDoublePendulum-v2")
    runner = agnes.Single(env, agnes.PPO, agnes.MLP)
    runner.log(agnes.log, agnes.TensorboardLogger(".logs/"), agnes.CsvLogger(".logs/"))
    runner.run()

agnes.log - object of StandardLogger class that outputs parameters to console. agnes.TensorboardLogger - class for writing logs in Tensorboard file. agnes.CsvLogger - class for writing logs in csv file. (required for plotting)

DistributedMPI

Unlike in Single runner, in DistributedMPI runner due to async executing, weights are delayed by one rollout but this has no effect on learning because weights are delayed only by one update as it is in Single runner. So all parameters like probabilities ratio stay the same.

Runs with

mpiexec -n 3 python -m mpi4py script_name.py
or
mpirun -n 3 python -m mpi4py script_name.py

This command will run 2 workers and 1 trainer.

# script_name.py
import agnes


if __name__ == '__main__':
    env = agnes.make_vec_env("BreakoutNoFrameskip-v4")
    runner = agnes.DistributedMPI(env, agnes.PPO, agnes.CNN)
    runner.run()

Algorithms

A2C

Sync version of Advantage Actor Critic is implemented in this framework and can be used simply:

import agnes


if __name__ == '__main__':
    runner = agnes.Single(env, agnes.A2C, agnes.MLP)
    runner.run()

PPO

Proximal Policy Optimization is implemented in this framework and can be used simply:

import agnes


if __name__ == '__main__':
    runner = agnes.Single(env, agnes.PPO, agnes.MLP)
    runner.run()

Neural Network Architectures

Multi Layer Perceptron

Can be used with both continuous and discrete action spaces.

...
runner = agnes.Single(env, agnes.PPO, agnes.MLP)
...

Convolutional Neural Network

Can be used only with discrete action spaces.

...
runner = agnes.Single(env, agnes.PPO, agnes.CNN)
...

Recurrent Neural Network

Can be used with both continuous and discrete action spaces.

...
runner = agnes.Single(env, agnes.PPO, agnes.RNN)
...

Convolutional Recurrent Neural Network

Can be used only with discrete action spaces.

...
runner = agnes.Single(env, agnes.PPO, agnes.RNNCNN)
...

Convolutional Neural Network with last LSTM layer

Can be used only with discrete action spaces.

...
runner = agnes.Single(env, agnes.PPO, agnes.LSTMCNN)
...

Make environment

• make_vec_env(env, envs_num=ncpu, config=None)**

  Parameters:

  • env(str or function) is id of gym environment or function, that returns initialized environment
  • envs_num(int) is a number of environments to initialize, by default is a number of logical cores on the CPU
  • config(dict) is a dictionary with parameters for Monitor and for initializing environment, by default is None(uses default config)

  Returns:

  • dict of
    1. "env"(VecEnv object)
    2. "env_type"(str)
    3. "env_num"(int) is a number of envs in VecEnv object
    4. "env_name"(str) is the name of envs in VecEnv object(Id in gym or class name)

  The whole tuple should be put in a runner.

• make_env(env, config=None) is an alias of make_vec_env without envs_num argument that will be setted to 1.

Notice: Some plot functions and environment wrappers were taken from OpenAI Baselines(2017).