Hi, I’m looking for a CL benchmark that would allow for agents with simple action-perception loops. Ideally this would be VERY simple, because I’m just getting started and I don’t want to get bogged down in perception problems. I guess something like CL for RL which can run on a single machine in minutes/hours. Any pointers would be much appreciated, thanks!
Check out these two papers, they list very interesting envs (mostly low-dimensional) that have been already used for CL:
- [2012.13490] Towards Continual Reinforcement Learning: A Review and Perspectives (arxiv.org)
- [1905.10112] Continual Reinforcement Learning in 3D Non-stationary Environments (arxiv.org)
The new version of Avalanche (v0.2.0) will natively support the construction of CL benchmarks based on gym environments. We are looking how to integrate stable-baselines within Avalanche for the future.
Finally I’d really love to explore the NetHack / MiniHack Learning environment which has very simple tasks as well as very complex ones: very cool for prototype small & scale fast: