Dataset Supporting the Manuscript: Staged NEAT Integration for Risk-Sensitive Expected-Value Decisions

This dataset supports the research manuscript "Staged NEAT Integration for Risk-Sensitive Expected-Value Decisions". ### Research Hypothesis In noisy and stochastic environments, decision-making models must learn non-linear probability-payoff representations (such as product, ratio, or fractional-power relationships). While trainable connection-level exponent gates offer a strong inductive bias for such structures, simultaneously evolving network topology, weights, and exponents (joint evolution) under finite-sample stochastically noisy rewards results in severe training instability due to search space explosion. We hypothesize that a decoupled, staged neuroevolutionary framework ("Staged NEAT Integration")—which first searches for and freezes a compact topology (Phase 1, b_topo) and subsequently optimizes connection-level exponent parameters (Phase 2, b_exp)—stabilizes the search trajectory, yielding superior expected-value (EV) decision alignment and minimal decision regret. ### What the Data Shows & Notable Findings This database represents a massive experimental campaign consisting of 1,440 independent training runs, 145,440 decision rows, and 31,680 population-dynamics rows: 1. Primary Decoupled Search Matrix: Demonstrates that under a fixed computational budget of B=2048, the optimal MLP-16 staged configuration (b_topo=128, b_exp=1920) achieves an exceptional 97.75% expected-value alignment and a negligible 0.064% regret, outperforming joint evolution by +5.750 percentage points. The optimal MLP-32 and MLP-64 staged configurations achieve 97.00% and 97.38% EV-alignment respectively, improving over joint evolution by +3.438 and +3.312 percentage points. 2. Robustness Extensions: Logs evolutionary trajectories across diverse payoff distributions (Bernoulli Jackpot, Lognormal Tail, Pareto Tail, Two-Point Moderate), loss-framed tasks, and nonstationary environments (featuring mid-run expected-value-ratio shifts). 3. Real-World Decision Generalization: Validates the framework across temporal UCI Bike-Sharing data (17,379 records), 101,766 clinical hospital encounters (Diabetes 130-US), and 30,000 credit card default profiles under five asymmetric domain utility rules (Expected Value, Risk-Averse, Loss-Averse, Critical-Miss Weighted, Profit-Cost). The results demonstrate robust, statistically significant improvements in domain-specific decision utilities. ### How to Interpret and Use the Data Researchers can leverage this database to: - Study how allocating training budgets between topology discovery (b_topo) and parameter tuning (b_exp) shifts the Pareto frontier of neural networks. - Evaluate the learning trajectories and speciation dynamics of neuroevolutionary algorithms under noisy, non-additive stochastic rewards. - Verify the paired difference statistical significance (utility delta, bootstrap 95% confidence intervals, and p-values) to benchmark novel decision models.本数据集支持研究论文《面向风险敏感期望价值决策的分阶段NEAT集成（Staged NEAT Integration）》。 ### 研究假设 在含噪随机环境中，决策模型需学习非线性概率-收益表征（如乘积、比值或分数幂关系）。可训练的连接级指数门控为这类结构提供了较强的归纳偏置，但在有限样本随机噪声奖励下，同时演化网络拓扑、权重与指数（联合演化）会因搜索空间爆炸引发严重的训练不稳定。我们提出的分阶段神经演化框架（“分阶段NEAT集成”）——先搜索并冻结紧凑拓扑结构（阶段1，b_topo），随后优化连接级指数参数（阶段2，b_exp）——能够稳定搜索轨迹，实现更优异的期望价值（EV）决策对齐度与最低决策遗憾。 ### 数据集呈现结果与重要发现 本数据库包含一项大规模实验研究，共计1440次独立训练运行、145440条决策记录与31680条种群动力学记录： 1. **主解耦搜索矩阵**：在固定计算预算B=2048下，最优多层感知机（MLP）-16分阶段配置（b_topo=128，b_exp=1920）实现了97.75%的期望价值对齐度与仅0.064%的决策遗憾，相较联合演化提升5.750个百分点。最优MLP-32与MLP-64分阶段配置的期望价值对齐度分别为97.00%与97.38%，相较联合演化分别提升3.438与3.312个百分点。 2. **鲁棒性拓展实验**：记录了多种收益分布下的演化轨迹，包括伯努利大奖分布、对数正态尾部分布、帕累托尾部分布、双点适中分布，损失框架任务，以及包含运行中期望价值比值偏移的非平稳环境。 3. **真实世界决策泛化性验证**：在多个真实数据集上验证了该框架的有效性，包括时序UCI共享单车数据集（17379条记录）、101766例临床医院就诊记录（美国130家医院糖尿病数据集），以及5种非对称领域效用规则下的30000条信用卡违约档案（效用规则包括期望价值、风险规避、损失规避、关键失误加权、收益-成本加权）。实验结果表明，该框架在领域特定决策效用上实现了稳健且具有统计显著性的提升。 ### 数据集解读与使用方式 研究人员可借助本数据库开展以下研究： - 探究拓扑发现（b_topo）与参数调优（b_exp）之间的训练预算分配如何改变神经网络的帕累托前沿。 - 评估含噪、非加性随机奖励下神经演化算法的学习轨迹与种群分化动力学。 - 验证配对差异统计显著性（效用增量、自助法95%置信区间与p值），以基准测试新型决策模型。