Search Algorithms for Computing Stock Composition of a Mixture from Traits ofIndividuals by Maximum Likelihood

摘要

The conditional maximum likelihood method of estimating stock-mixture compositionis described for discrete characters. Computer programs were developed for several general-purpose, nonlinear optimization algorithms, specialized to searching for the conditional maximum likelihood estimate (CMLE); and their performances were compared for hypothetical and real-world stock mixtures. Measures of performance were search time, failure rate, and stability of CMLE distributions as the criterion for stopping search (guaranteed percent achieved of the maximum of the likelihood function or GPA) was increased. Programs based on the conjugate gradient (with square root transform of stock composition) and expectation maximization algorithms were superior in reliability and speed. Iteratively-reweighted least squares programs produced the most stable CMLE distributions because their terminal GPAs typically exceeded that specified by more than other programs.