手法を比較

	多腕バンディット（UCB、トンプソンサンプリング）	適応的臨床試験デザイン	逐次 / 群逐次試験デザイン
分野	実験計画法	実験計画法	実験計画法
系統	Hypothesis test	Hypothesis test	Hypothesis test
提唱年	1952	1994	1979
提唱者	Robbins (1952); UCB1 by Auer et al. (2002); Thompson sampling by Thompson (1933)	Bauer & Köhne	O'Brien & Fleming; Pocock; Lan & DeMets
種類	Sequential decision / bandit algorithm	Adaptive hypothesis test with interim analyses	Adaptive stopping trial design
原典	Auer, P., Cesa-Bianchi, N., & Fischer, P. (2002). Finite-Time Analysis of the Multiarmed Bandit Problem. Machine Learning, 47(2–3), 235–256.	Bauer, P. & Köhne, K. (1994). Evaluation of Experiments with Adaptive Interim Analyses. Biometrics, 50(4), 1029–1041.	O'Brien, P.C. & Fleming, T.R. (1979). A Multiple Testing Procedure for Clinical Trials. Biometrics, 35(3), 549–556.
別名	MAB, bandit algorithm, UCB1, Thompson sampling	adaptive design, group sequential design, sample size re-estimation, platform trial	group sequential design, adaptive stopping design, Ardışık Deneme Tasarımı (Sequential / Group Sequential)
関連	4	3	3
概要	The multi-armed bandit (MAB) is an adaptive experimental framework that allocates trials sequentially across competing arms to minimise cumulative regret while simultaneously learning which arm performs best. Formalised by Robbins in 1952 and given finite-time guarantees by Auer et al. (2002), it balances exploration of uncertain options against exploitation of currently known best options — outperforming classical A/B testing whenever early stopping or cost-sensitive allocation matters.	Adaptive clinical trial design is a flexible experimental framework, formalised by Bauer and Köhne in 1994, in which pre-specified rules allow the trial to be modified mid-course — adjusting sample size, treatment arms, or randomisation ratios — based on accumulating interim data while rigorously controlling the Type I error rate.	Sequential and group sequential trial designs allow a study to be stopped early — or continued — based on interim analyses conducted as data accumulate. The core framework was formalised by O'Brien and Fleming in 1979 and extended by Lan and DeMets's alpha-spending approach, and it controls the overall Type I error rate across all planned looks by pre-specifying both efficacy and futility boundaries before enrolment begins.