Notes
What I took away from the papers on my reading list.
2026
- DAgger: imitation learning as no-regret online learning
Behaviour cloning trains a policy with supervised learning on states the expert visited. The trouble is that once the learner acts, it visits states the expert never did, and its mistakes there compound: an error early in the episode moves the learner further off-distribution, which makes the next error more likely. The paper makes this…