A collection of lifelong/continual learning baselines for reinforcement learning.
This repository is a growing collection of lifelong/continual learning baselines for reinforcement learning. If you’ve ever tried to track down and run baselines from the literature for a paper, you know how tedious and frustrating it can be. The goal here is to make that process faster and less error-prone by providing library of implementations of published baselines built on the Modulating Masks/DeepRL codebase and modified if needed to specifically experiment on reinforcement learning benchmarks. This way you can run find relevant algorithms in a familiar environment and compare results quicker and more easily.
Inspired by the Avalanche continual learning repository for supervised learning. https://github.com/ContinualAI/continual-learning-baselines and CORA https://github.com/AGI-Labs/continual_rl
If you implement additional algorithms from the literature in the Modulating Masks/DeepRL codebase, please contribute them to this repository as a new branch. This project only works if we maintain it together. Don't forget to also update the table below:
- EWC, Elastic Weight Consolidation (multihead) (description) (cite)
- SI, Synaptic Intelligence (multihead) (description) (cite)
- CLHNET, Continual Learning with Hypernetworks (description) (cite)
- CLEAR, Continual Learning with Experience Replay (description) (cite)
- Mask LC/BLC/RI, Modulating Masks (description) (cite)
- Mask SC, Mask Selective Combination (description) (cite)
TODO: Create a proper table for the above information