Panel Discussion
summary
Shows that regret‑optimal bandit algorithms lose precision on low arms but a quick practical fix is to inflate CIs by 1.5 so UCB/Thompson sampling become near‑useful for inference; simulations use ~1,000 samples and live runs aim for 1,500–2,000 observations per category with online updates after an initial batch, but non‑stationarity and contextual imbalance often make random data beat adaptive for contextual policy evaluation. Developmental work finds kids do both more random and more directed exploration and play-as-simulation helps train robust agents.
AI-written summary · about