This repo contains implementations of the algorithms, architectures, and environments from Duan et al., 2016 - 'RL^2: Fast Reinforcement Learning via Slow Reinforcement Learning', and Mishra et al., 2017 - 'A Simple Neural Attentive Meta-Learner'.

It has also recently been redesigned to facilitate rapid prototyping of new stateful architectures for memory-based meta-reinforcement learning agents.

The main idea of RL^2 is that a reinforcement learning agent with memory can be trained on a distribution of environments, and can thereby learn an algorithm to effectively transition from exploring these environments to exploiting them.

In fact, the RL^2 training curriculum effectively trains an agent to behave as if it possesses a probabilistic model of the possible environments it is acting in.

This theoretical background of RL^2 is discussed by Ortega et al., 2019 and a concise treatment can be found in my blog post.

Install the following system dependencies:

sudo apt-get update
sudo apt-get install -y cmake openmpi-bin openmpi-doc libopenmpi-dev

Installation of the system packages on Mac requires Homebrew. With Homebrew installed, run the following:

brew install cmake openmpi

Once the system dependencies have been installed, it's time to install the python dependencies. Install the conda package manager from https://docs.conda.io/en/latest/miniconda.html

Then run

conda create --name pytorch_rl2 python=3.8.1
conda activate pytorch_rl2
git clone https://github.com/lucaslingle/pytorch_rl2
cd pytorch_rl2
pip install -e .

To train the default settings, you can simply type:

mpirun -np 8 python -m train

This will launch 8 parallel processes, each running the train.py script. These processes each generate meta-episodes separately and then synchronously train on the collected experience in a data-parallel manner, with gradient information and model parameters synchronized across processes using mpi4py.

To see additional configuration options, you can simply type python train.py --help. Among other options, we support various architectures including GRU, LSTM, SNAIL, and Transformer models.

By default, checkpoints are saved to ./checkpoints/defaults. To pick a different checkpoint directory during training, you can set the --checkpoint_dir flag, and to pick a different checkpoint name, you can set the --model_name flag.

Our implementations closely matched or slightly exceeded the published performance of RL^2 GRU (Duan et al., 2016) and RL^2 SNAIL (Mishra et al., 2017) in every setting we tested.

In the tables below, n is the number of episodes per meta-episode, and k is the number of actions. Following Duan et al., 2016 and Mishra et al., 2017, in our tabular MDP experiments, all MDPs have 10 states and 5 actions, and the episode length is 10.

To perform policy optimization, we used PPO. We used layer norm instead of weight norm, and we report peak performance over training. Our performance statistics are averaged over 1000 meta-episodes.

In all cases, for training we used a configuration where the total number of observations per policy improvement phase was equal to 240,000. This is comparable to the 250,000 used in prior works. The per-process batch size was 60 trajectories. There were 8 processes. There were 8 PPO optimization epochs per policy improvement phase.

All other hyperparameters were set to their default values in the train.py script, except for the SNAIL experiments, where we used --num_features=32 due to the skip-connections.