This repository provides the code for the Model-Agnostic Meta-Learning (MAML) paper, focusing on few-shot reinforcement learning experiments. It enables researchers and practitioners to quickly adapt deep learning models to new tasks with minimal data, a key challenge in RL.

How It Works

MAML is a meta-learning algorithm that learns a model's initial parameters such that it can be rapidly fine-tuned for new tasks. The core idea is to optimize for a sensitive initialization that performs well after a few gradient steps on a new task, using a second-order meta-gradient update. This approach allows for fast adaptation without requiring task-specific architectures.

Quick Start & Requirements

• Install: Based on the rllab framework. Follow rllab installation instructions.
• Dependencies: TensorFlow v1.0+. Python 3.5+ recommended for rllab.
• Environments: Includes pointmass and MuJoCo environments.
• Documentation: https://rllab.readthedocs.org/en/latest/

Highlighted Details

• Implements Model-Agnostic Meta-Learning for fast adaptation in RL.
• Code accompanies the paper "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks" (Finn et al., ICML 2017).
• Includes scripts for running experiments from the paper.

Maintenance & Community

• Developed by UC Berkeley and OpenAI.
• For questions or issues, open an issue on the GitHub tracker.

Licensing & Compatibility

• The rllab framework, which this code is based on, is compatible with OpenAI Gym.
• Licensing details for maml_rl itself are not explicitly stated in the README, but rllab's licensing should be considered.

Limitations & Caveats

The code is noted as being particularly slow, with suggestions for contributions to improve parallelization and graph computation speed.