Chapter Four · failure evidence

What Multi-Objective Optimization got wrong, from 52 dissertations

Multi-objective optimization methods face challenges including computational bottlenecks, solver convergence failures, and distortion from scalarization weights. Practitioners frequently encounter situations where simpler baselines, sequential approaches, or dedicated single-objective formulations outperform complex Pareto algorithms. These records come from PhD theses at 20 institutions, 2021 to 2026. Each links to its thesis. They were extracted by language models reading the full text, so treat each as a lead to read, not a verdict.

Weighted sum scalarization and objective ratios distort the search landscape and introduce weight selection bias

10 theses · 9 institutions

Linear combinations, ratio formulations, and extreme objective weights created distorted search spaces, poor solver conditioning, and inferior trade-offs. Researchers rejected weighted-sum and scalarized methods in favor of Pareto front or epsilon-constraint approaches to eliminate subjective weighting bias and fixed front assumptions.

Tried and failed

unclamped linear combination multi-objective fitness function applied to multi-target molecular property optimization. Outcome: worse than baseline. Reason: unbounded objective contributions distorted the search landscape, severely reducing multi-target success rates

Design of (De)Polymerizable Polymers Using Machine Learning-Based Predictive Models and Generative Algorithms · Georgia Tech

Tried and failed

directly optimizing ratio of objectives applied to multi-objective trade-off optimization. Outcome: worse than baseline. Reason: optimizing quality-to-cost ratio degraded the Pareto frontier and reduced outcome quality versus optimizing quality alone

Essays on developing artificial intelligence solutions for patient-centered healthcare delivery · UT Austin

Considered and rejected

Considered and rejected: Rejected scalarized single-objective fixed trade-off penalty as primary method because it requires arbitrary weighting and fails to properly explore the multi-objective Pareto trade-off frontier.

Fairness-Aware Unsupervised Learning Methods for Healthcare Applications · Penn

Tried and failed

Single-objective extreme weighting in multi-objective heuristic search applied to constrained spatial path finding. Outcome: worse than baseline. Reason: Extremes degrade search efficiency and Pareto trade-offs compared to balanced joint optimization.

Semantically diverse and spatially constrained queries · Iowa State

Tried and failed

extreme single-objective weighting in multi-objective optimization applied to coupled multidisciplinary systems design. Outcome: did not converge. Reason: extreme objective weights create severe gradient scaling imbalances and poor conditioning for gradient-based solvers

An Integrated Vehicle, Payload, and Trajectory Optimization Framework for Highly-Coupled Aircraft Systems · MIT

Tried and failed

weighted-sum multi-objective optimization for spatio-temporal trade-offs applied to remote sensing trajectory and payload design. Reason: Optimizer favored spatial averaging over multi-pass temporal averaging under the specified resolution objectives

An Integrated Vehicle, Payload, and Trajectory Optimization Framework for Highly-Coupled Aircraft Systems · MIT

Considered and rejected

Considered and rejected: Rejected a single weighted-sum multi-objective function in favor of generating a Pareto frontier via the constraint method due to varying stakeholder priorities and hidden objectives.

Expanding Geographic Access to Fertility Care in the United States: A Statistical and Multiobjective Optimization Approach · JScholarship

Considered and rejected

Considered and rejected: Rejected auxiliary weighted sum functions for multi-objective optimization because they impose fixed Pareto front assumptions

Evolution of Multiobjective Neuromodulated Neurocontrollers for Multi-Robot Systems · Carleton University Institutional Repository

Considered and rejected

Considered and rejected: Rejected weighted-sum multi-objective optimization to avoid arbitrary a priori weights and dimensionless transformation bias.

Development Of A Methodology For Fast Optimization Of Building Retrofit And Decision Making Support · Penn

Considered and rejected

Considered and rejected: Rejected using lexicographic selection or weighted-sum multi-objective optimization for COVID-19 scoring in favor of Pareto optimization with NSGA-II followed by post-hoc selection based on prognostic agreement.

Evolutionary Optimization of Decision Trees for Interpretable Reinforcement Learning · IRIS - UNITN - prod

Considered and rejected

Considered and rejected: Rejected standard weighted sum multi-objective approach in favor of epsilon-constraint combined with TOPSIS to eliminate subjective weight-selection bias.

Enhancing Thermoelectric Generator Performance · Scholarship at UWindsor Institutional Repository

Standard multi-objective evolutionary and metaheuristic algorithms underperform simpler baselines or hybrid methods

9 theses · 9 institutions

Algorithms like NSGA-II, particle swarm optimization, and genetic algorithms suffered from low convergence speed, particle aggregation, or poor diversity compared to alternatives like HMOPSO and dynamic programming. Experimental benchmarks demonstrated that candidate evolutionary and heuristic methods were beaten by simpler designs, alternative genetic methods, or specialized baseline tools.

Lost to a baseline

For multi-objective soft valve optimization, Taguchi experimental design found a 4-valve Pareto front whereas Bayesian Optimization produced only 1 Pareto-optimal design.

Elastomeric Strain Limitation for Design of Soft Pneumatic Actuators · Penn

Lost to a baseline

GA Method 4 test MAE on multi-objective optimization (0.067) was beaten by GA Method 2 (0.0655) and GA Method 1 (0.0658).

Data-Driven Discovery of Molecular Catalysts for Liquid Organic Hydrogen Carriers · DSpace at SUNY Buffalo

Lost to a baseline

For memory overhead optimization in multi-objective tuning, Particle Swarm Optimization (PSO) underperformed Genetic Algorithm (GA) (inv_overhead score of 83% vs. 92%)

Resource-Efficient Edge Computing and Lightweight Traffic Fingerprinting for Scientific Applications · unevada

Tried and failed

multi-objective genetic algorithm optimization applied to satellite constellation configuration design. Outcome: worse than baseline. Reason: Poor Pareto front diversity caused the algorithm to miss valid multi-satellite solutions and produce inferior coverage metrics.

Satellite Constellation Optimization for In-Situ Sampling and Reconstruction of Tides in the Thermospheric Gap · Virginia Tech

Tried and failed

ensemble Pareto evolutionary algorithm applied to multi-objective combinatorial core optimization. Outcome: worse than baseline. Reason: Ensemble approach failed to outperform standalone evolutionary strategy across benchmark scenarios.

Light Water Reactor Loading Pattern Optimization with Reinforcement Learning Algorithms · MIT

Lost to a baseline

For time overhead optimization in multi-objective tuning, Artificial Bee Colony (ABC) achieved the lowest score compared to GA and PSO

Resource-Efficient Edge Computing and Lightweight Traffic Fingerprinting for Scientific Applications · unevada

Lost to a baseline

NSGA-II and MOPSO lost to HMOPSO on multi-objective benchmark functions (ZDT1, ZDT2, ZDT3, ZDT6), where NSGA-II showed low convergence speed and MOPSO suffered from particle aggregation/collapse.

Numerical simulation and optimization of fracturing-stimulated reservoir volume · Texas Tech

Considered and rejected

Considered and rejected: Rejected the Multi-Objective Evolutionary Algorithms NSGA-II, Borg, and SAMODE in favor of ε-NSGAII because they lacked approximate Pareto frontiers or used unsuitable dominance functions.

Ship and Naval Technology Trades-Offs for Science And Technology Investment Purposes · Georgia Tech

Lost to a baseline

Beam Search (BS) was outperformed by Multi-Objective Dynamic Programming (MODP) in retaining solution diversity across delta-v bands on a 1026-asteroid catalog.

Trajectory design of multi-target missions via graph transcription and dynamic programming. · Cranfield

Lost to a baseline

MOMoT beat MDEOptimiser on multi-objective REF-EXT-A (median HV 0.53 vs 0.46)

Model-Driven Optimization with a Focus on the Effectiveness and Efficiency of Evolutionary Algorithms · open_UMR Marburg DSpace 10.0

High computational overhead and function evaluation requirements make finding full Pareto fronts impractical

8 theses · 7 institutions

Heuristic searches, multi-objective reinforcement learning, and Bayesian optimization were rejected because sampling across entire preference spaces required excessive evaluations and GPU resources. Generating full Pareto frontiers or performing exact geometric Boolean operations imposed prohibitive runtimes without necessarily improving solution quality.

Considered and rejected

Considered and rejected: Rejected relying purely on standard heuristic algorithms (GA, PSO, SA) for multi-objective resonator tuning due to excessive function evaluations and curse of dimensionality.

Optimization of Locally Resonant Metafoundations for the Protection of Industrial Tanks and Small Modular Reactors Subjected to Low-Frequency Seismic Waves · IRIS - UNITN - prod

Considered and rejected

Considered and rejected: Rejected multi-objective RLHF (Reinforcement Learning from Human Feedback) due to training instability, extensive hyperparameter tuning requirements, and computational inefficiency compared to MODPO.

Essays on Digital Content Strategies: Creation, Diffusion, and Monetization · Harvard

Considered and rejected

Considered and rejected: rejected multi-objective Bayesian optimization (MOBO) due to requirements of sampling across all preferences and inability to efficiently update distributions sequentially

Preserving Correct Behaviors during Neural Network-based Control Policy Repair · Penn

Considered and rejected

Considered and rejected: Rejected Multi-Objective Genetic Algorithm (MOGA) and ε-constraint optimization due to high computational overhead.

Optimum performance controls of multiport converter for transport electrification applications · University of Nottingham Repository

Tried and failed

exact constructive solid geometry Boolean union operations applied to multi-objective geometric shape optimization. Outcome: too slow. Reason: computational overhead increased runtime nearly threefold without improving the Pareto hypervolume solution quality

Early Phase Performance Driven Design Assistance Using Generative Models · Harvard

Considered and rejected

Considered and rejected: A posteriori multi-objective preference articulation rejected for real-time MPC due to high computational effort of executing multiple optimization runs with different weights.

Energy management for industrial plants with thermal batch processes Energiemanagement für Industrieanlagen mit thermischen Batch-Prozessen · DSpace-CRIS at TU Wien

Considered and rejected

Considered and rejected: Rejected Pareto front multi-objective optimization approaches due to prohibitive computational costs in finding and representing the entire Pareto front.

Bridging the gaps in trustworthy autonomy: Alignment, validation, and physical execution in learning-enabled cyber-physical systems · Iowa State

Considered and rejected

Considered and rejected: Rejected multi-policy/separate network architectures for multi-objective RL due to high GPU/computational training burden in stable-baselines

Light Water Reactor Loading Pattern Optimization with Reinforcement Learning Algorithms · MIT

Multi-objective optimizers fail to converge due to local optima entrapment and algorithmic under-parameterization

7 theses · 6 institutions

Gradient-based solvers, simulated annealing with simultaneous high-dimensional perturbations, and Bayesian optimization frequently trapped in local minima across complex Pareto landscapes. Default fixed genetic operator parameters and acquisition functions with dimensionality mismatch further prevented algorithms from converging to valid Pareto solutions.

Tried and failed

simultaneous perturbation of all variables in simulated annealing applied to multi-objective system optimization. Outcome: did not converge. Reason: acting like ineffective random search, high-dimensional perturbations failed to find improved solutions

Multi-Objective System Optimization of a Mars Atmospheric ISRU Plant · MIT

Tried and failed

NSGA-II with default genetic operator parameters applied to multi-objective tri-level network optimization. Outcome: did not converge. Reason: fixed genetic crossover/mutation probability parameters hindered optimization convergence without adaptive tuning and offsets

Resilience-driven planning and operation of networked microgrids featuring decentralisation and flexibility · Imperial

Tried and failed

Gaussian process Bayesian optimization with weighted expected improvement applied to multi-objective circuit parameter optimization. Outcome: did not converge. Reason: trapped in local optima and failed to satisfy multiple stringent constrained specifications simultaneously

Efficient optimization methods for analog/mixed-signal integrated circuits via machine learning · UT Austin

Considered and rejected

Considered and rejected: Rejected direct search and gradient-based optimization algorithms because direct search lacks precision for multi-objective trade-offs and gradient methods are prone to trapping in local extrema

DIGITAL CODESIGN OF A POWER MODULE WITH INTEGRATED THERMAL MANAGEMENT · Georgia Tech

Tried and failed

L-BFGS-B gradient-based optimization applied to multi-objective quantum circuit parameter tuning. Outcome: did not converge. Reason: Trapped in local minima across the complex multi-objective optimization landscape.

Classical and quantum optimization of quantum processors · Harvard

Tried and failed

multi-objective Bayesian optimization with L-BFGS acquisition applied to high-dimensional benchmark functions. Outcome: did not converge. Reason: acquisition function optimization became under-parameterized when objectives approached input dimensionality

Multiscale Modeling and Microstructure Design of Biodegradable Polymeric Scaffolds · Georgia Tech

Tried and failed

Bayesian optimization with minimal initialization applied to multi-objective soft actuator design. Outcome: did not converge. Reason: Premature overconfidence and exploitation of a local minimum prevented discovering the true Pareto front.

Elastomeric Strain Limitation for Design of Soft Pneumatic Actuators · Penn

Multi-objective Pareto formulations underperform dedicated single-objective methods when single goals dominate

6 theses · 5 institutions

Pareto compromise solutions yielded inferior outcomes compared to dedicated single-objective allocations or manual exploration when individual targets were prioritized. Authors rejected multi-objective Pareto frontiers because stakeholders preferred converging on a single configuration or using single-objective weighted cost functions where metrics map to financial cost.

Tried and failed

multi-objective evolutionary optimization for single target priority applied to constrained resource allocation. Outcome: worse than baseline. Reason: Pareto compromise solutions underperform dedicated single-objective optimization when only one metric is prioritized

Integrating social equity into sustainable programming of road projects : a quantitative approach · UT Austin

Lost to a baseline

For individual single-objective scores (e.g., access metric f1 alone or precision f2 alone), single-objective adaptive allocations scored higher on their respective single metric than the multi-objective proposed method.

Operationalization of AI Systems Through Data-Centric Design · Georgia Tech

Lost to a baseline

ADD-MES and DPT-BO achieved smaller RX coil areas (19.26 mm2 and 11.04 mm2 vs 215.8 mm2) than the thesis's WPT multi-objective design

Scattering and Optimization Methods for Radar and Millimeter Wave Applications · Georgia Tech

Considered and rejected

Considered and rejected: Rejected multi-objective Pareto optimization for trading off task time and energy consumption, choosing a single-objective weighted cost function where both metrics map directly to financial cost.

Reinforcement Learning and Trajectory Optimization for the Concurrent Design of high-performance robotic systems · IRIS - UNITN - prod

Considered and rejected

Considered and rejected: Rejected multi-objective optimization yielding a Pareto front in favor of single weighted-sum fitness because users prefer converging on a single configuration solution.

Analyzing the impact of configurations in data-driven software applications · Iowa State

Tried and failed

multi-objective optimization generative tool in design process applied to architectural parametric structural design. Outcome: worse than baseline. Reason: algorithmic optimization tool led designers to produce lower quantitative performance than purely manual design exploration

Generative Design Tools: Implications on Design Process, Designer Behavior, and Design Outcomes · MIT

Mismatches between sequential parameter optimization and simultaneous multi-objective formulation degrade solution quality

5 theses · 5 institutions

Optimizing parameters sequentially or performing component substitution without joint re-optimization prematurely restricted the search space and drifted away from the Pareto front. Conversely, attempting simultaneous joint optimization across conflicting physical mechanisms or distinct parameter sets proved ineffective and prompted researchers to separate the problem into sequential steps.

Tried and failed

sequential multi-objective optimization applied to catalyst material candidate discovery. Outcome: worse than baseline. Reason: Optimizing properties sequentially prematurely restricts search space compared to simultaneous joint Hamiltonian optimization

Optimization Algorithms for Quantum and Digital Annealers · Harvard

Tried and failed

component substitution without joint parameter re-optimization applied to charged particle beam injector optimization. Outcome: worse than baseline. Reason: optimizing a single upstream parameter shifts the operating point away from the multi-objective Pareto front

Towards High Beam Brightness from Photocathodes · Cornell

Considered and rejected

Considered and rejected: Rejected multi-objective simultaneous optimization across multiple stress-strain curves in favor of a sequential single-objective optimization isolated to specific dislocation mechanisms.

Modeling of overloads in cyclic loading · Cranfield

Considered and rejected

Considered and rejected: Rejected using only separate single-objective Bayesian optimizations for multi-objective problems because candidate solutions along the trade-off front were too sparsely distributed.

A DATA-DRIVEN APPROACH FOR MICROSTRUCTURAL INVERSE DESIGN THROUGH MATERIALS INFORMATICS · JScholarship

Considered and rejected

Considered and rejected: Rejected merging calibration and retrofit multi-objective optimization into a single unified step due to distinct input parameter sets and conflicting objective functions

Rapid Workflow for Energy Model Calibration, Retrofit Optimization, and Uncertainty Analysis of Large Building Energy Retrofits · Carleton University Institutional Repository

Left open by the authors

Problems the authors named and did not get to.

Left open

Develop multi-objective optimization algorithms to design macromolecules balancing functionality, synthetic accessibility, cost, and carbon footprint. Blocker: Lack of specific objectives, target datasets, property models, or optimization formulation

Designing Macromolecules using Machine Learning and Simulations · MIT

Left open

Incorporate multi-objective cost trade-offs into multi-fidelity batch Bayesian optimisation algorithms. Blocker: Lacks specific formulation, mathematical approach, or evaluation target for multi-cost trade-offs

Towards improving Bayesian optimisation for the physical sciences via long-term planning · Imperial

Left open

Formulate and evaluate multi-objective functions for hydrogen-CCS network optimization considering resource consumption and energy security alongside cost. Blocker: No specific multi-objective formulation, metrics, or mathematical definitions are provided.

Modelling framework for the design of hydrogen-CCS networks to decarbonise heating and industrial clusters · Imperial

Left open

Develop and integrate a multi-objective optimization decision-support tool for Pond-In-Pond wastewater treatment operational strategies. Blocker: Optimization criteria, specific situational constraints, and integration architecture with the hydrodynamic models are not defined.

Water sustainability using the Pond-In-Pond treatment system with reuse. · Texas Tech

Left open

Analyze multi-objective submodular optimization with noisy evaluation oracles and approximately submodular objective functions. Blocker: No concrete approach, target bounds, or formal evaluation criteria are specified

Submodular optimization problems in the context of social influence · Iowa State

Left open

Implement multi-objective decision-making methods to balance security and economic trade-offs in risk-constrained optimal power flow. Blocker: The unfinished work lacks specific mathematical formulation or designated multi-objective algorithms to implement

Assessment of marginal value of transmission services based on risk assessment of system security · Iowa State

Left open

Test the iterative batch Bayesian Optimization and D-MPNN-guided genetic algorithm on multi-objective optimization combining synthesizability, off-target effects, and docking scores. Blocker: None

Molecular Property Predictors for Downstream De Novo Generation · Harvard

Left open

Implement multi-objective Bayesian optimization models to simultaneously predict salt solubility, chemical compatibility, and conductivity in polymer electrolytes. Blocker: Lacks experimental or computational dataset for salt solubility and chemical compatibility corresponding to these polymer electrolyte formulations

Polymer Electrolyte Discovery via Rational Design and High Throughput Methods · MIT

Left open

Develop and implement a multi-objective Bayesian optimization acquisition function that relies on a single Gaussian process model instead of separate models. Blocker: None

Multiscale Modeling and Microstructure Design of Biodegradable Polymeric Scaffolds · Georgia Tech

Left open

Extend the surrogate optimizer to multi-objective Pareto optimization incorporating carbon valuations alongside financial metrics in agent-based retrofit simulation. Blocker: None

Investigating the use of Machine Learning for Resource Intensive Agent-Based Simulation and Optimisation of Domestic Energy Retrofit Adoption · University of Nottingham Repository

Checking a claim in this area?

We can run the same search on any method or claim. If nothing turns up, we will say so, and that proves nothing on its own.