PREDICTING THE JOB PERFORMANCE BY USING MULTI-OBJECTIVE OPTIMIZATION AND GMDH-TYPE NEURAL NETWORK TO IMPROVE PERSONNEL SELECTION PROCESS

Failure or success of an organization is directly related to the way in which human resource is recruited. Personnel selection is a complex system whose objective is to assess the dierences among the candidates and to select the most appropriate person. Achieving this goal can be facilitated by predicting the future performance; however, cognitive limitations of the human mind make the behavior of unknown and/or very complex systems dicult to predict. Using current employee performance data to predict the future behavior of the applicants is an interesting area. Since personnel selection system is coupled with the ambiguity and uncertainty values, rstly, it is necessary to model the imprecise modes of reasoning to make rational decisions in an environment of uncertainty and imprecision. Secondly, optimal design of model parameters based on the existing data should be considered in order to minimize errors and to maximize adaptability of the model. In this regard, the present study identies eective input variables in predicting the output after designing the personnel selection system, whose output is job performance and its dimensions. The \Emotional Quotient (EQ)" and \individual variables" are also selected as input variables. Then, input and output data are collected from operational personnel work in Guilan Gas Company. Next, an approach that utilizes Genetic Algorithms (GAs) is applied for multi-objective design of group method data handling (GMDH)-type neural networks in order to model the job performance. In this algorithm, training error (TE) and prediction error (PE) are simultaneously minimized. However, the nature of human resource is lled up with uncertainty and inherent ambiguity, the correlation coecient is 0.9956 and RMSEA is just 0.06 which indicate high accuracy of extracted model and the maximum adaptability to predict job performance with actual performance. It is worth mentioning that the extracted models of performance dimensions are eective for 84% to 96% of the data, and the performance variable is exactly the same as the real value. Therefore, the presented model is able to receive some information as inputs and to predict the future performance which has the minimum error. Accordingly, the most eective input variables in predicting performance are also optimally determined.