Irreversible adaptive allocation rules (Q581980)

Real valued populations \(\Pi_ 1,...,\Pi_ k\) are given, with \(\Pi_ i\) having density \(f_ i(x| \theta)\) with respect to some fixed measure, \(i=1,...,k\). The unknown parameter \(\theta\) lies in an interval of the real line. Also given are reward functions \(g_ i(x,\theta)\). A sample of size N is taken sequentially, and at the n-th stage an observation \(X_ n\) is obtained from \(\Pi_ i\), for some choice of i. The selection is made with help of a sequence \(\{\phi_ n\), \(n=1,...,N\}\) of choice functions, where each \(\phi_ n\) has possible values 1,...,k and may depend on the choices and observed X's through stage n-1. Furthermore, the restriction \(\phi_ n\leq \phi_{n+1}\) is imposed (irreversible allocation rule), which arises naturally in certain experiments, e.g., in sequential bioassays. Several additional assumptions are made. The objective is to maximize the expected reward \(J_ N(\theta)=\sum^{N}_{n=1}E_{\theta}g_{\phi_ n}(X_ n,\theta)\) by optimum choice of the sequence \(\phi_ n\). Since this depends on the unknown \(\theta\), an adaptive rule is proposed. Several asymptotic (as \(N\to \infty)\) optimum properties are proved.