scientific article; zbMATH DE number 3947447
[For the entire collection see Zbl 0542.00012.] Let the real valued random variable X have the density f(x,\(\theta)\) with respect to a measure \(\nu\) on the real line R, with \(\theta\in \Theta \subset R\). It is assumed that \(\mu(\theta)\equiv E_{\theta}X<\infty\) for every \(\theta\in \Theta\). For \(\theta,\lambda\in \Theta\) define \[ I(\theta,\lambda)=E_{\theta}(\log f(X,\theta)/\log f(X,\lambda)). \] Let \(\pi_ 1,...,\pi_ k\) be k populations (not necessarily distinct) such that \(\pi_ j\) has density \(f(x,\theta_ j)\) with respect to \(\nu\), with \(\theta_ j\) belonging to \(\Theta\) but otherwise unknown. It is assumed that the numbers \(\mu_ j=\mu (\theta_ j)\) are not all equal. Put \(\mu^*=\max \{\mu_ j:\) \(j=1,...,k\}.\) At the \(n^{th}\) stage of sampling, one of the populations is chosen and an observation \(x_ n\) from this population is obtained. The choice of population may depend on \(x_ 1,...,x_{n-1}\) and on previous choices. The object is to maximize the expected value of \(S_ n\equiv x_ 1+...+x_ n\) asymptotically for large n. If the (or a) population with largest mean were known, then continued sampling from this population would be optimal and E \(S_ n=n \mu^*\). In the absence of this knowledge an optimal allocation rule minimizes the regret \(R_ n=n \mu^*-E S_ n\) in an asymptotic sense. It is shown that if \(\mu(\lambda)>\mu^*\) for some \(\lambda\in \Theta\), then a lower bound for \(\liminf_{n\to \infty} R_ n/\log n\) is \(\sum (\mu^*-\mu_ j)/I(\theta_ j,\lambda)\), where the summation is over all j for which \(\mu_ j<\mu^*\). If \(\sup_ n R_ n<\infty\) the regret is said to be bounded. Two types of problems are treated. In the first of these, there is given a number \(\gamma\) that separates \(\mu^*\) from the next smallest value of \(\mu_ j\), say \(\lambda\). Either \(\lambda <\gamma <\mu^*\), or \(\gamma =\mu^*\), or \(\gamma =\lambda\). In each of the three cases an allocation rule is given for which the regret is bounded. In the second type of problem f is assumed to share some of the properties of a one-parameter exponential family; in particular that \(\mu\) (\(\theta)\) is increasing. Define \(\theta^*=\max \theta_ j\) (so that \(\mu (\theta^*)=\mu^*)\) and \({\bar \theta}=\inf \{\lambda \in \Theta:\) \(\lambda >\theta^*\}\) (\({\bar \theta}=\infty\) if there is no such \(\lambda)\). An allocation rule is proposed for which it is shown that the regret is bounded if \({\bar \theta}=\infty\), and if \({\bar \theta}<\infty\) then \[ \lim R_ n/\log n=\sum (\mu^*-\mu_ j)/I(\theta_ j,{\bar \theta}) \] so that the rule is optimal.
- Asymptotically efficient adaptive allocation rules
- Optimality of random allocation design for the control of accidental bias in sequential experiments
- Sequential allocation to two treatments that converges to an - superior treatment
- A control problem arising in the sequential design of experiments
- Adaptive treatment allocation and the multi-armed bandit problem
- Certainty equivalence control with forcing: Revisited
- Information-regret compromise in covariate-adaptive treatment allocation
- Large deviation asymptotics for statistical treatment rules
- Arbitrary side observations in bandit problems
- scientific article; zbMATH DE number 3846699 (Why is no real title available?)
- scientific article; zbMATH DE number 410134 (Why is no real title available?)
- scientific article; zbMATH DE number 1031931 (Why is no real title available?)
- On large deviations properties of sequential allocation problems
- Game of thrones: fully distributed learning for multiplayer bandits
- Optimal Alpha Spending for Sequential Analysis with Binomial Data
- Implementing Optimal Allocation in Sequential Binary Response Experiments
- Irreversible adaptive allocation rules
- Functional Sequential Treatment Allocation
- Optimal Bayesian strategies for the infinite-armed Bernoulli bandit
This page was built for publication:
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3718014)