TEST 4

This dataset supports the paper "Optimizing Expert Knowledge Investments in the Software Industry: A Markov Decision Process Approach to Innovation in Knowledge Management" (submitted to Journal of Innovation & Knowledge, Elsevier).
The dataset contains the complete results of a three-round modified Delphi expert elicitation study conducted to calibrate the transition probabilities and model parameters of a Markov Decision Process (MDP) framework for expert knowledge management. Thirteen domain experts (5 Chief Technology Officers and 8 Senior Engineering Managers) from the Argentine software industry participated across three rounds over six weeks (combined experience: 147 years; M = 11.3, SD = 2.9).
The file comprises seven sheets: (1) expert panel demographics and selection criteria; (2) the complete Delphi questionnaire instrument covering 47 parameters across six categories (transition probabilities, reward function coefficients, fatigue model, knowledge decay rates, cost structure, and discretization thresholds); (3–5) individual expert responses for Rounds 1, 2, and 3, with convergence statistics per parameter (median, mean, SD, IQR, CV); (6) convergence analysis demonstrating full consensus by Round 3 — Mean IQR converged from 0.28 to 0.12 (threshold < 0.15), the count of probability parameters with IQR > 0.20 decreased from 23 to 0 out of 31 (threshold < 5), and Mean CV decreased from 0.41 to 0.18 (threshold < 0.25); and (7) final adopted parameter values with full traceability to the MDP transition function and reward specification. Round 3 mean values are adopted uniformly as the primary central tendency estimator for all MDP parameters (transition probabilities P1–P20, action costs P34–P38, reward coefficients P21–P24, and the fatigue multiplier threshold P26), motivated by the distributional properties of the expert responses — particularly the right-skewed distribution of parameters such as α₃ (CV = 0.687), where the mean provides a more informative estimate than the median. Likert-scale validation parameters P44–P47 constitute a single declared exception, adopting the operational integer value by scale constraint. The Round 3 median column is retained for reference and sensitivity analysis; a full means-vs-medians sensitivity is reported in Section 4.6 of the manuscript.
This dataset enables complete reproducibility of all model parameters from raw expert judgment to final adopted values (47 parameters × 13 experts × 3 rounds = 1,833 data points).

本数据集为已提交至爱思唯尔（Elsevier）旗下《创新与知识》（Journal of Innovation & Knowledge）期刊的论文《优化软件行业的专家知识投入：知识管理创新的马尔可夫决策过程方法》提供数据支撑。
本数据集包含一项三轮改进德尔菲法专家征询研究的完整成果，该研究用于校准面向专家知识管理的马尔可夫决策过程（Markov Decision Process, MDP）框架的转移概率与模型参数。来自阿根廷软件行业的13名领域专家（含5名首席技术官与8名高级工程经理）参与了为期六周的三轮调研，其总从业年限为147年，均值（M）为11.3年，标准差（SD）为2.9。
该数据集文件包含7个工作表：(1) 专家小组人口统计学特征与遴选标准；(2) 完整的德尔菲法调查问卷工具，涵盖6大类共47项参数，具体包括转移概率、奖励函数系数、疲劳模型、知识衰减率、成本结构与离散化阈值；(3–5) 第1、2、3轮的专家个体应答数据，附带每项参数的收敛统计量，包括中位数、均值、标准差（Standard Deviation, SD）、四分位距（Interquartile Range, IQR）与变异系数（Coefficient of Variation, CV）；(6) 收敛性分析结果，证明第3轮已达成完全共识——平均四分位距从0.28收敛至0.12（阈值<0.15），31项概率参数中四分位距大于0.20的数量从23降至0（阈值<5），平均变异系数从0.41降至0.18（阈值<0.25）；(7) 最终采用的参数值，且可完整追溯至MDP转移函数与奖励规范。
研究统一采用第3轮的均值作为所有MDP参数的主要集中趋势估计量，涵盖转移概率P1~P20、动作成本P34~P38、奖励系数P21~P24与疲劳乘数阈值P26。这一选择基于专家应答的分布特性——尤其是α₃等参数的右偏分布（变异系数CV=0.687），此时均值相比中位数能提供更具信息量的估计值。李克特量表（Likert-scale）验证参数P44~P47为唯一例外，根据量表约束采用操作整数值。第3轮的中位数列仍予以保留，用于参考与敏感性分析；完整的均值-中位数敏感性分析结果已在论文的4.6章节中报告。
本数据集可实现从原始专家判断到最终采用参数的所有模型参数的完全可复现性，共计47项参数×13名专家×3轮调研=1833个数据点。