Avalanche RL is a fork of ContinualAI's Pytorch-based framework Avalanche with the goal of extending its capabilities to Continual Reinforcement Learning (CRL), bootstrapping from the work done on Super/Unsupervised Continual Learning.

It should support all environments sharing the gym.Env interface, handle stream of experiences, provide strategies for RL algorithms and enable fast prototyping through an extremely flexible and customizable API.

The core structure and design principles of Avalanche are to remain untouched to easen the learning curve for all continual learning practitioners, so we still work with the same modules you can find in avl:

• Benchmarks for managing data and stream of data.
• Training for model training making use of extensible strategies.
• Evaluation to evaluate the agent on consistent metrics.
• Extras for general utils and building blocks.
• Models contains commonly used model architectures.
• Logging for logging metrics during training/evaluation.

Head over to Avalanche Website to learn more if these concepts sound unfamiliar to you!

Features

Features added so far in this fork can be summarized and grouped by module.

Benchmarks

RLScenario introduces a Benchmark for RL which augments each experience with an 'Environment' (defined through OpenAI gym.Env interface) effectively implementing a "stream of environments" with which the agent can interact to generate data and learn from that interaction during each experience. This concept models the way experiences in the supervised CL context are translated to CRL, moving away from the concept of Dataset toward a dynamic interaction through which data is generated.

RL Benchmark Generators allow to build these streams of experiences seamlessly, supporting:

• Any sequence of gym.Env environments through gym_benchmark_generator, which returns a RLScenario from a list of environments ids (e.g. ["CartPole-v1", "MountainCar-v0", ..]) with access to a train and test stream just like in Avalanche. It also supports sampling a random number of environments if you wanna get wild with your experiments.
• Atari 2600 games through atari_benchmark_generator, taking care of common Wrappers (e.g. frame stacking) for these environments to get you started even more quickly.
• Habitat, more on this later.

Training

RLBaseStrategy is the super-class of all RL algorithms, augmenting BaseStrategy with RL specific callbacks while still making use of all major features such as plugins, logging and callbacks. Inspired by the amazing stable-baselines-3, it supports both on and off-policy algorithms under a common API defined as a 'rollouts phase' (data gathering) followed by an 'update phase', whose specifics are implemented by subclasses (RL algorithms).

Algorithms are added to the framework by subclassing RLBaseStrategy and implementing specific callbacks. You can check out this implementation of A2C in under 50 lines of actual code including the update step and the action sampling mechanism. Currently only A2C and DQN+DoubleDQN algorithms have been implemented, including various other "utils" such as Replay Buffer.

Training with multiple agent is supported through VectorizedEnv, leveraging Ray for parallel and potentially distributed execution of multiple environment interactions.

Evaluation

New metrics have been added to keep track of rewards, episodes length and any kind of scalar value (such as Epsilon Greedy 'eps') during experiments. Metrics are kept track of using a moving averaged window, useful for smoothing out fluctuations and recording standard deviation and max values reached.

Extras

Several common environment Wrappers are also kept here as we encourage the use of this pattern to suit environments output to your needs. We also provide common gym control environments which have been "parametrized" so you can tweak values such as force and gravity to help out in testing new ideas in a fast and reliable way on well known testbeds. These environments are available by pre-pending a C to the env id as in CCartPole-v1 as they're registered on first import.

Models

In this module you can find an implementation of both MLPs and CNNs for deep-q learning and actor-critic approaches, adapted from popular papers such as "Human-level Control Through Deep Reinforcement Learning" and "Overcoming catastrophic forgetting in neural networks" to learn directly from pixels or states.

Logging

A Tqdm-based interactive logger has been added to ease readability as well as sensible default loggers for RL algorithms.

Quick Example

import torch
from torch.optim import Adam
from avalanche.benchmarks.generators.rl_benchmark_generators import gym_benchmark_generator

from avalanche.models.actor_critic import ActorCriticMLP
from avalanche.training.strategies.reinforcement_learning import A2CStrategy

# Config
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# Model
model = ActorCriticMLP(num_inputs=4, num_actions=2, actor_hidden_sizes=1024, critic_hidden_sizes=1024)

# CRL Benchmark Creation
scenario = gym_benchmark_generator(['CartPole-v1'], n_experiences=1, n_parallel_envs=1, 
    eval_envs=['CartPole-v1'])

# Prepare for training & testing
optimizer = Adam(model.parameters(), lr=1e-4)

# Reinforcement Learning strategy
strategy = A2CStrategy(model, optimizer, per_experience_steps=10000, max_steps_per_rollout=5, 
    device=device, eval_every=1000, eval_episodes=10)

# train and test loop
results = []
for experience in scenario.train_stream:
    strategy.train(experience)
    results.append(strategy.eval(scenario.test_stream))

Compare it with vanilla Avalanche snippet!

Check out more examples here (advanced ones coming soon) or in unit tests. We also got a small-scale reproduction of the original EWC paper (Deepmind) experiments.

Installation

As this fork is still under development, the advised way to install it is to simply clone this repo git clone https://github.com/NickLucche/avalanche.git and then just follow avalanche guide to install as developer. Spoiler, just run conda env update --file environment-dev.yml to update your current environment with avalanche-rl dependencies. Currently, the only added dependency is ray.

Disclaimer

This fork is under strict development so expect changes on the main branch on a fairly regular basis. As Avalanche itself it's still in its early Alpha versions, it's only fair to say that Avalanche RL is in super-duper pre-Alpha.

We believe there's lots of room for improvements and tweaking but at the same time there's much that can be offered to the growing community of continual learning practitioners approaching reinforcement learning by allowing to perform experiments under a common framework with a well-defined structure.