项目验收的同行评议辅助决策评价方法研究——基于贝叶斯正则化修正的BP人工神经网络模型

Research on Method to Assist the Decision-making of Project Evaluation through Peer Review——Based on BP Artificial Neural Network Model with Bayesian Regularization Correction

[目的/意义]个体的同行评议判断不可避免地会由于专家主观性导致评审结果嵌入式有偏,而个体成员聚集在一起开展有组织的群体决策时,由于客观组织评审标准的存在,能有效避免个体非理性因素导致的系统整体偏差。[方法/过程]为了找到更适合的定量化模型来模拟专家组群体的复杂评审决策,本研究借助BP人工神经网络模型,样本选取2001-2005年间批准实施的林业"948计划"项目126项,从"投入—产出"视角选取表征项目特征的36个指标,基于以上指标通过有监督的机器学习来模拟项目评审专家组(5~9名成员)的决策打分值,并添加贝叶斯正则化修正项来提高模型的预测精度。[结果/结论]所建构的同行评议打分预测模型在添加贝叶斯正则化修正项后,平均误差平方和由10-3上升到10-4数量级,而模型的预测值与真实值间相关系数ρ由0.37(0.33)上升到0.61(0.47),模型实现了在个体水平上对项目评审打分的较准确预测,即通过精准计量评价来有效辅助同行评议决策。

[ Purpose/Significance] The peer-review judgment of the individual will inevitably lead to the embedded bias of the result of review due to subjectivity. However,individual members gather together to carry out organized group decision-making,which can effec-tively avoid the overall systematic deviation caused by individual irrationality due to the existence of objective criteria of review. [ Method/ Process]In order to find a more suitable quantitative model to simulate the complex decision-making of the group of experts,this paper u-ses the BP artificial neural network model to select 126 items of forestry ＂ 948 projects＂ approved from 2001 to 2005 through “ input-out-put＂ perspective to characterize the projects. Basing on the above characteristics through the supervised machine learning,this paper simu-lates the project review results of decision-making of expert group （5-9 members） and adds Bayesian Regularization correction to improve the prediction accuracy of the model. [ Results/Conclusions] The average squared errors of the model is increased from 10-3 to 10-4 after the addition of the Bayesian Regularization term. The correlation coefficient p between the predicted value and the true value is increased from 0. 37 （0. 33） to 0.61 （0.47）. The model achieves the accurate prediction of the project score at the individual level and assists thepeer review decision-making by metrics.