OHRID-Net: Many-objective hyperparameter reduction in deep networks using fuzzy-preference sorting, with application to deceased-donor kidney allocation

Abstract The selection of deep network hyperparameters and topology remains largely manual, and evolutionary architecture search usually optimizes a single accuracy objective. When one model design must support several conflicting prediction tasks while remaining compact, the search becomes a many-objective problem: most candidates are mutually non-dominated, and Pareto ranking alone cannot select among them. We propose many- O bjective H yperparameter R eduction I n D eep Net works (OHRID-Net), an evolutionary framework with two components. A fixed-length binary encoding represents the hyperparameters chosen for tuning, together with a free layer-connection topology, as one chromosome whose length follows from the search space, and which always decodes to a connected network. A fuzzy-preference sorting operator replaces crowding-distance and reference-point selection: membership functions and if–then rules convert each candidate's objectives into one preference score, so stakeholders state priorities in their own terms. Applied to national deceased-donor records from the Organ Procurement and Transplantation Network (2016–2022), OHRID-Net jointly optimized hard-to-place kidney identification, provisional acceptance, primary acceptance and trainable-parameter count. Elite networks reached median accuracies of 87.3%, 86.1% and 89.9%, with best values within 0.3–1.7 percentage points of separately optimized single-task networks, while using 8.1–24.3% of the worst-case parameter budget. In a single run, two added stakeholder rules moved the primary-acceptance distribution in the intended direction. On public benchmark problems, fuzzy sorting converged faster and more tightly than standard non-dominated sorting genetic algorithms while focusing on a preferred region, and priority rules steered it reliably toward the prioritized objective.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-13
DOI
https://doi.org/10.21203/rs.3.rs-11517203/v1
Primary Topic
Advanced Multi-Objective Optimization Algorithms
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

OHRID-Net: Many-objective hyperparameter reduction in deep networks using fuzzy-preference sorting, with application to deceased-donor kidney allocation

Ci̇han H. Dağli, Lirim Ashiku
1 links
Research Square
Advanced Multi-Objective Optimization Algorithms
preprint

OHRID-Net: Many-objective hyperparameter reduction in deep networks using fuzzy-preference sorting, with application to deceased-donor kidney allocation

Ci̇han H. Dağli, Lirim Ashiku
preprint en

Abstract

Abstract The selection of deep network hyperparameters and topology remains largely manual, and evolutionary architecture search usually optimizes a single accuracy objective. When one model design must support several conflicting prediction tasks while remaining compact, the search becomes a many-objective problem: most candidates are mutually non-dominated, and Pareto ranking alone cannot select among them. We propose many- O bjective H yperparameter R eduction I n D eep Net works (OHRID-Net), an evolutionary framework with two components. A fixed-length binary encoding represents the hyperparameters chosen for tuning, together with a free layer-connection topology, as one chromosome whose length follows from the search space, and which always decodes to a connected network. A fuzzy-preference sorting operator replaces crowding-distance and reference-point selection: membership functions and if–then rules convert each candidate's objectives into one preference score, so stakeholders state priorities in their own terms. Applied to national deceased-donor records from the Organ Procurement and Transplantation Network (2016–2022), OHRID-Net jointly optimized hard-to-place kidney identification, provisional acceptance, primary acceptance and trainable-parameter count. Elite networks reached median accuracies of 87.3%, 86.1% and 89.9%, with best values within 0.3–1.7 percentage points of separately optimized single-task networks, while using 8.1–24.3% of the worst-case parameter budget. In a single run, two added stakeholder rules moved the primary-acceptance distribution in the intended direction. On public benchmark problems, fuzzy sorting converged faster and more tightly than standard non-dominated sorting genetic algorithms while focusing on a preferred region, and priority rules steered it reliably toward the prioritized objective.

Research Square
Missouri University of Science and Technology (US)
Advanced Multi-Objective Optimization Algorithms
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon