The Application Study Of Artificial Neural Networks In Evaluating Basic Research Achievement Of Science And Technology

ObjectiveTo construct a comprehensive evaluating system of basic scientific and technologic achievements in which multiple indexes should be comprised, the input and output data be dealt with methods of scientometrics , mathematics and statistics , and systemic differentiation be implemented with artificial neural network ( Multilayer Perceptions as network type and Conjugate Gradient Decent as training algorithm).MethodData about 256 projects of nation - wide and multidiscipline were collected from a ministry of China as example cases. They were recommended to be awarded "natural science prize" concerning basic research achievements. The evaluating indexes (input variables) included project source, duration of performance, publication of research articles, monographies produced, patents obtained , achievements communicated on meetings, articles collected by retrieving systems, articles cited by research articles, articles cited by monographies in foreign languages, meeting articles cited and monographies cited. The output variable was the prize - rank. The input variables were dealt with mathematically and assigned randomly and appropriately into three sets - training sets, verification sets and test sets according to their prize - ranks and disciplines in turn. Data were firstly trained in STATISTICA Neural Networks (SNN) by Multilayer Perceptron and Conjugate Gradient Decent till rational and ideal network resulted. Secondly, the data were simulated with Logistic regression ( multinomial logistic and binary logistic). Note: there were two ways in evaluation. One wasone - step evaluation, i.e. all prize - ranks were yielded in one step. The other was two - step evaluation, that is, all the projects were divided into two groups (awarded or not) before the awarded ones determined their prize - rank. In each way/step, neural nework and logistic regression were applied respectively.ResultThe accordance rates of the results of three simulation methods in two evaluating ways with the outcome of peer review resulted. See the following table;Table 1 Accordance rates of SNN, Logistic Regression with peer review (% )The indexes screened. And models of evaluation established.Input (independent) variables included in the models are as followings; all the ,11 variables in Neural Network; articles cited, duration of project, monogra-phies cited, and meeting articles (p<0.15) in Logistic models.ConclusionIn terms of accordance, two - step way excelled one - step; that in neural network excelled than that with logistic. In terms of variables included, neural network was far more of capacity than logistic regression. We can concluded that in dealing with such nonlinear problems as evaluation of scientific and technologic achievement, neural network precedes nonlinear statistical methods. Evaluating results of neural network can be viewed as an important and available reference when scientific and technologic achievements are judged by peer review.