Estimating sample sizes and statistical filtering of on-line measurements

Two important features which affect the efficiency of an on-line optimization strategy are noise and the time spent on measuring the optimization criterion, or let it be called sample size. The time spent on measuring the optimization criterion, i.e. the sample size, may slow down the control strategy. Consequently the sample size will result from a compromise and be small enough not to slow down the strategy unduly, and large enough to estimate the true value of the sample with confidence. In this paper a simple statistical technique is suggested to solve both problems. Nevertheless, classical filtering techniques adapted to computers do exist; in this paper a second order filter will be considered for comparison. However, the statistical solution may be favoured when implementing it in an on-line control strategy: it follows closely the current stability of the process, has almost no time delay and is more akin to an algorithm requiring very little computation. (4 refs).