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

Published: 14 August 2026| Version 2 | DOI: 10.17632/jvwzmn8jpt.2
Contributors:
,
,

Description

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.

Files

Steps to reproduce

To reproduce the expected-value decision datasets, population dynamics logs, and real-world clinical and financial utility evaluations, follow this workflow: 1. Dependencies and Environment: Set up a Python 3.12 environment with PyTorch, scikit-learn, NumPy, and a modified NEAT-python package customized to support decoupled, staged training phases. 2. Staged Neuroevolution Execution (Synthetic Tasks): a. Initialize the base neural network backbones (MLP-16, MLP-32, or MLP-64) and hold the total budget B (512 or 2048 option-pairs) fixed. b. Execute Phase 1 (Topology Search): Run the evolutionary algorithm using b_topo option-pairs to discover a compact network structure. c. Handoff & Freeze: Select the fittest topology in the population and freeze its structural connections. d. Execute Phase 2 (Exponent Tuning): Evolve the connection-level exponent parameters using the remaining b_exp budget (B - b_topo). Test both Phase 2 Mode A (weights frozen, exponents evolve) and Mode B (weights and exponents evolve together). e. Complete 1,440 independent training runs across 32 random seeds to populate the 145,440 decision rows and 31,680 population-dynamics rows. 3. Robustness Extensions: Run population-level evolution across Bernoulli Jackpot, Lognormal Tail, Pareto Tail, and Two-Point Moderate payoff distributions, as well as loss-framed (Bernoulli and Lognormal Shock) and nonstationary tasks (featuring mid-run expected-value-ratio shifts). 4. Real-World Dataset Preprocessing and Generalization: a. Download the public datasets from the UCI Machine Learning Repository: Bike-Sharing Dataset (17,379 records), Diabetes 130-US Hospitals (101,766 clinical encounters), and Default of Credit Card Clients (30,000 financial profiles). b. Apply domain-specific utility functions (Expected Value, Risk-Averse, Loss-Averse, Critical-Miss Weighted, Profit/Cost) to capture asymmetric decision costs. c. Split the data chronologically (training on prior periods, testing on subsequent periods) and log final decision accuracy, utility, AUC, and false negative rates. 5. Export: Collect all run-by-run metrics and paired utility differences into structured CSV files for paired t-test significance testing.

Categories

Artificial Intelligence, Machine Learning

Funders

Licence