Three-alternative learning rating system for decision support system of process automation

The problem of information synthesis of learning decision support system for scintillator crystals growth from the melt is under consideration. As a criterion of functional efficiency modification of Kullback - Leibler information measure is proposed: Dkl(p,q) = X xe _z(P(x) _ q(x))ln^ , q(x) where p(x) and q(x) are the probability functions of two sets of discrete random value x e R, and D _KL is the distance between the ensembles {p} and {q}. To improve the functional efficiency of decision support systems (DSS), the ideas and methods of information and intellectual extreme technology were used [5-9], based on maximizing the information ability of the system in the learning process. The multimodal three-alternative learning rating system is proposed as more appropriate for this process. The formula of the criteria is ₁ K (k) + _K (k) + _K (k) E ) = (2(K ₁ ) + K? + K ))/n _min _3) log ₂- 1 +K +K ₍k) , 8 1 2 3 min (3n _min _K _K2 _K3 ) where K h K^, and K 3 are respectively the number of events, which define belonging recognizable implementations to the classes «LESS THAN NORMAL», «NORMAL», «MORE THAN NORMAL», if they are really belong to the training data matrixes classes on the k step of learning; nmin is the minimum amount of representative training sample. This approach provides the opportunity to work with three classes in the same recognition feature space, monitoring the radiuses of corresponding containers. This improves the quality of the recognition system. Proposed criteria for three-alternative system allows to evaluate the functional efficiency of the DSS with random distributions of recognition implementations in the feature space, and thus, has the versatility compared to the unimodal classifier with the same learning parameters [12]. In addition, the use of the MFE to build learning rules characterized by higher reliability and efficiency in comparison with two-alternative criterion. This increase in efficiency of learning algorithm is achieved due to the absence in the process of its implementation averaging the criteria and the simultaneous construction of decision rules for three classes.

Keywords

Authors

References