A new approach for increasing the accuracy of radiometric measurement systems by using the most informative collections of pairs of stochastically coupled signals is described for the example of gamma/beta distributed signals. The criterion of informativeness of a collection of signals is substantiated for an arbitrary distribution law for n signals. A method is presented for selecting an informative collection of signals and for the learning process. The algorithm for determining the values of the parameter is based on probability distances. To increase the efficiency of technological processes in industrial production it is necessary to have reliable information about many parameters. Such information is obtained with the aid of subsurface radiometric measurement systems. The difficulty of obtaining reliable information is due to the complicated conditions of the technological processes and the noisiness of the signals. For this reason it is impossible to determine with acceptable error the controllable parameter g on the basis of one signal. This led to expanded use of the results of high information fluxes and not individual measurements. To increase accuracy it is often necessary to resort to corrections. We have shown that the possibility of correction is strongly limited, and in most cases in practice correction even decreases the accuracy of the result [1]. For this reason, corrections should be introduced only if three conditions are satisfied simultaneously: possibility, desirability, and necessity. The accuracy of the result is also increased by measuring several signals and determining according to them the required parameter using increasingly more complicated signal processing algorithms. However, even the best algorithms result only in a decrease in the amount of information contained in the collection of signals employed. For this reason, an effective measure for increasing accuracy is to select the most informative signals from the entire initial set of signals. Ordinarily, the optimal choice of signals is made independently of the quality of the measuring scheme [2]. It is observed that the recognition reliability tends to increase when measurements of several signals are performed and all modified algorithms for simultaneous processing of the signals are used. Since the amount of information decreases with processing, the importance of selecting the signals increases sharply [2, 3]. In [4] a one-dimensional/3- and y-distribution is described. In [5] a two-dimensional y-,/3-distribution is investigated. The aim of the present article is to show that the accuracy of radiometric systems can be sharply increased by using pairs of signals correlated with one another. Essence of the New Approach. The conventional signal-selection methods are based on maximum entropy for a chosen collection of signals. Independent signals ensure maximum entropy. At first glance this seems obvious - why measure a second signal, if it can be calculated on the basis of the stochastic coupling with the first signal. As a result, the selected signals turned out to be independent. However, analysis of exchange of information in nature shows that a small number of strongly correlated indications (signals) is always used for recognition, since even the human brain can operate simultaneously with not more than seven signals [6, 7]. In nature, recognition is improved by selection of the smallest number of indicators which are closely coupled with one another, i.e., by means of learning. The fact that technology and nature develop in opposite directions from the standpoint of recognition stimulated the present author to analyze this paradox, especially since the idea of using dependent signals to increase recognition reliability is well known [8], though it has not been further developed correctly. Analysis of the paradox showed that nature is right, since under favorable conditions a collection of specially correlated signals contains much more information than the like number of stochastically independent signals with the same variance.