Position: Deployed Reinforcement Learning should be Continual

In three linesPosition paper arguing deployed RL systems should adopt continual learning instead of train-then-fix paradigm. Authors identify four sources of post-deployment non-stationarity requiring never-ending learning and analyze real-world continual RL examples.

Summary generated by Claude — human-verified