Quality Diversity for Reliable Data Driven Time-Use Optimization

AI summary

Overview

• Research area: Machine learning for health behavior — specifically evolutionary quality diversity (QD) optimization combined with uncertainty quantification, applied to compositional time-use (24-hour activity) data.
• Technical level: Intermediate. Readers should be comfortable with regression models, statistical variance formulas, and the general idea of evolutionary/archive-based optimization, but the paper explains the health and optimization background in accessible terms.
• Scope: A single-conference-paper-style preprint (arXiv:2608.05230v1 [stat.ML], 05 Aug 2026) that introduces an uncertainty-aware MAP-Elites framework for time-use recommendations and tests it on 4D and 7D real-world child cohort health models.

What This Paper Is About

Predictive models can estimate how someone's 24-hour activity composition relates to health outcomes like body mass index, life satisfaction, and cognition, but those models are uncertain — they make shaky predictions in regions of the data they have rarely seen. Standard time-use optimization only maximizes predicted benefit, which can push recommendations into poorly supported regions of the model and produce unrealistic advice. This paper builds uncertainty directly into the optimization so that recommended daily activity compositions balance expected health benefit against the model's confidence in that prediction.

Key Contributions

1. An uncertainty-quantified Quality Diversity framework for time-use optimization. The authors extend MAP-Elites for time-use (previously applied without uncertainty) so that each candidate solution is scored by a risk-aware fitness, F(x) = E(x) − C_α · U(x) for maximization (and F(x) = E(x) + C_α · U(x) for minimization), where E(x) is the expected health benefit and U(x) the prediction uncertainty.
2. A derivation of prediction uncertainty from a regression model over compositional time-use data. Starting from multiple linear regression on isometric log-ratio (ilr) transformed time-use variables, they derive U*(x₀) = σ̂ · sqrt(1 + x₀ᵀ(XᵀX)⁻¹x₀), showing that uncertainty depends only on where a solution sits relative to the predictor data, not on the predicted value itself.
3. Exploration of two behavioral representations under uncertainty. They study variable-based behavioral spaces (VBS, descriptors taken from solution variables such as LPA and MVPA) and objective-based behavioral spaces (OBS, descriptors taken from objective outputs such as LS and cognition), showing how uncertainty shapes each map differently.
4. Comprehensive real-data experiments on 4D and 7D health models using a large child cohort (n > 1000), covering objectives such as VO2, life satisfaction, BMI z-score, cognition, academic performance, body fat percentage, and psychiatric score.

Main Findings

• Ignoring uncertainty matters. The authors report that standard (non-discounted) optimization selects high-benefit solutions in weakly supported model regions, so uncertainty is not uniform across the solution space and can change which solutions look best.
• Uncertainty effects are objective-specific. Three regimes emerge in the 4D VBS experiments: low-uncertainty reliable optima (VO2, where high-quality solutions align with low uncertainty), uncertainty-driven high-performance regions (LS and BMIz, where high-quality solutions are linked to higher uncertainty), and low-discriminative objectives (COG, where performance varies little and uncertainty mainly breaks ties).
• 4D benchmark numbers (standard approach, VBS over LPA and MVPA). VO2 elite [514, 450, 261, 215] with E(x) = 60.0145 and U(x) = 8.9677; LS elite [613, 436, 230, 161] with E(x) = 13215.0981 and U(x) = 4648.4160; BMIz elite [557, 536, 197, 150] with E(x) = 0.00001 and U(x) = 1.0958; COG elite [527, 738, 159, 16] with E(x) = 2.4523 and U(x) = 0.1365.
• Discounting barely changes the 4D maps. Because the 4D space has few behavioral variables, the standard and discounted maps stay largely similar, with only marginal local shifts in a few areas.
• Objective-based maps reveal relationships between health outcomes. In the 4D OBS, regions with higher cognitive scores and higher life satisfaction correspond to higher BMIz values, and the best-performing solution in that space has E(x) = 0.9626 with U(x) = 1.0968 at position (20, 10); the best cognitive value is E(x) = 2.4771 with U(x) = 0.1372 at position (1, 20). There are no solutions where cognitive and life satisfaction values are simultaneously very low or very high.
• In the 7D problem, discounting reduces uncertainty substantially. High-quality solutions concentrate in specific regions rather than being evenly distributed. For academics, the elite [480, 1, 28, 95, 137, 609, 90] reaches E(x) = 580.03 with U(x) = 73.41; for body fat, the elite [510, 281, 24, 284, 1, 320, 20] reaches E(x) = 3.22 with U(x) = 1.53. Under discounting, the academic elite reaches E(x) = 572.27 with U(x) = 65.08, and the body fat elite reaches E(x) = 3.07 (the uncertainty value is cut off in the available text). The authors note the color scales for the two approaches differ because of this reduction in uncertainty.
• Uncertainty is an indicator, not a uniform penalty. The authors conclude that uncertainty in QD reshapes the archive differently for each objective while still preserving behavioral diversity.

Methodology in Plain English

• Model the health outcomes. The team uses data from the Child Health CheckPoint — Growing Up in Australia study, a cross-sectional module in Release 9.1 C1 of the Longitudinal Study of Australian Children (LSAC). The LSAC birth cohort began in 2003 with over 10,000 participants; those remaining through Wave 9 (n = 7,658) were invited to the CheckPoint at ages 11–12. Daily behaviors came from 7-day, 24-hour wrist-worn accelerometry (4D composition: sleep, sedentary, LPA, MVPA) and 24-hour recalls (7D composition including domains such as screen time and school activities). Outcomes included NAPLAN academic performance, NIH Toolbox cognition, VO2 max via cycle ergometry, and standardized BMI z-scores.
• Handle the constant-sum constraint. Because activities must sum to 1440 minutes, the variables are linearly dependent. The authors apply Compositional Data Analysis and Simplexity via the compositions package in R, using a log-ratio expectation maximization algorithm for zero replacement followed by an isometric log-ratio (ilr) transformation, mapping the D-part simplex into D − 1 independent coordinates. A multiple linear regression is then fit, optionally including quadratic and interaction terms of the ilr coordinates if statistically significant.
• Compute uncertainty. From the regression model they derive the prediction uncertainty as a scaled quadratic form involving the (XᵀX)⁻¹ matrix and the new predictor vector, with σ estimated by the standard error of the regression. An example (XᵀX)⁻¹ matrix and s_e = 4634.088 are given for the 4D Life Satisfaction objective.
• Optimize with MAP-Elites plus a risk penalty. The algorithm (Algorithm 1) starts with G random solutions and then generates new candidates by mutation or crossover of archive members. Each candidate is placed into a behavioral-space cell via its descriptor, its E(x) and U(x) are computed, and its risk-aware fitness F(x) is compared against the incumbent elite in that cell. A solution replaces the elite only if it has better fitness. The archive, fitness, expected values, and uncertainties are all returned. Crossover and mutation operators are the same as in prior time-use QD work.
• Choose the penalty. The coefficient C_α = 6.36134 comes from the α-fractile point of the standard normal distribution at α = 1 − 10⁻¹⁰. The authors note uncertainty is not necessarily normally distributed in this way, and that differences in expected benefit usually dwarf differences in uncertainty, so this factor has a relatively moderate effect. Setting C_α = 0 recovers the standard model that ignores uncertainty.
• Run experiments. They compare standard and discounted approaches across 10 independent runs, terminating after 10⁶ fitness evaluations, for the 4D and 7D problems, in both VBS and OBS configurations.

Why This Matters

• Research impact. The paper connects uncertainty quantification with quality diversity optimization in a data-driven decision context, showing that QD archives can be used as maps of model reliability rather than only maps of performance. It also produces explicit, uncertainty-aware relationships between health outcomes, which pure maximization approaches obscure.
• Real-world applications.
  • Public health guidance: generating 24-hour movement recommendations that are robust rather than artifacts of extrapolation (relevant to guidelines such as the 2019 Australian 24-Hour Movement Guidelines, which specify 9–11 hours of sleep and ≥ 1 hour of MVPA for children).
  • Personalized behavioral counseling: the OBS maps let a clinician find the next-best composition if a person cannot, for example, increase MVPA, while showing how reliable that alternative prediction is.
  • Risk-aware intervention design: shifting recommendations toward lower-uncertainty regions when high-benefit solutions rest on weak model support.
  • Academic and cognitive health planning: understanding trade-offs, such as higher cognitive scores and life satisfaction co-occurring with higher BMIz in the OBS map.
• Industry relevance. The framework generalizes beyond health — the authors state the QD approach can be used for any model from which E(x) and U(x) can be derived — making it relevant to any organization deploying ML recommendations where predictions are extrapolative, such as scheduling, resource allocation, or consumer behavior guidance.

Future Directions

• Extension beyond linear regression. The authors note their model is based on multiple linear regression but that the QD approach works for any model providing E(x) and U(x); applying it to nonlinear or nonparametric predictors with different uncertainty estimates remains open.
• Better-calibrated uncertainty. The C_α = 6.36134 factor is borrowed from a normal distribution while uncertainty is not necessarily normally distributed, implying a need for more principled discounting or calibration.
• Higher-dimensional and differently structured behavioral spaces. The 4D setting was too constrained for discounting to change maps much, and the 7D analysis was restricted to 2D behavioral spaces to ensure convergence, leaving higher-dimensional descriptors unexplored.
• Completed 7D analysis and richer trade-off characterization. The available text cuts off mid-report of the discounted 7D results, and the paper's stated aim of systematically analyzing how uncertainty propagates across high-dimensional behavioral spaces leaves room for fuller study.

Target Audience

Researchers and practitioners at the intersection of evolutionary optimization, uncertainty quantification, and behavioral health — particularly those working on time-use epidemiology, compositional data analysis, or quality diversity algorithms. Health professionals and data scientists designing personalized or population-level activity recommendations will benefit from the applied sections, while ML researchers will find the QD-with-uncertainty formulation and the reliability-versus-benefit trade-off analysis most relevant.

Authors’ abstract

The daily allocation of the finite 24-hour time budget is strongly associated with physical, mental, and cognitive health. While predictive models can estimate the relationship between time-use compositions and health outcomes such as body mass index, life satisfaction, and cognition, most optimization approaches focus only on maximizing expected benefit and do not consider the uncertainty inherent in data-driven prediction. Ignoring uncertainty in health-related decisions can lead to unrealistic time-use recommendations. To address this gap, we introduce an uncertainty quantification Quality Diversity (QD) framework for a more reliable time-use recommendation. Objective functions are derived using compositional data analysis using a large child cohort dataset n > 1000, to capture the relationship between daily activity compositions and multiple health indicators. We develop a new approach that incorporates predictive uncertainty into QD processes and produces more reliable recommendations that balance the expected health benefits with the confidence of the model. We explore the solution space through variable-based and objective-based behavioral representations, revealing diverse high-quality time-use composition and explicit relationships between health outcomes under uncertainty. By embedding uncertainty directly into optimization, our framework shifts the time-use recommendations toward regions of lower uncertainty while preserving high-quality structures for more reliable decision-making in behavioral health.