Chapter Four · failure evidence
What Optimal Resource Allocation & Scheduling got wrong, from 71 dissertations
The records detail common failure patterns encountered when designing and applying algorithms for optimal resource allocation and scheduling across computing, manufacturing, and network domains. Researchers repeatedly face trade-offs where exact formulations become computationally intractable, simple heuristics violate complex system constraints, and learning-based or static policies fail under dynamic operating conditions. 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.
Greedy and localized heuristics cause resource fragmentation and violate global constraints
Heuristic dispatching and greedy assignment strategies make myopic placement decisions that ignore cumulative downstream bottlenecks, non-convex return thresholds, or bandwidth and buffer limits. These localized choices lead to severe resource fragmentation, system deadlock, and degraded performance compared to global optimization.
Tried and failed
greedy heuristics for resource allocation applied to dynamic model serving under SLOs. Outcome: worse than baseline. Reason: greedy optimization fails to achieve the optimal Pareto trade-off between accuracy and latency constraints
Efficient AI Stack: Deployment-Aware Neural Architecture Search and Serving of Deep Neural Networks · Georgia Tech
Tried and failed
graph centrality heuristics for DAG partitioning applied to distributed workflow scheduling. Outcome: worse than baseline. Reason: failed to account for downstream cumulative resource bottlenecks
Leveraging Feedback for Dynamic Execution Optimization · Penn
Tried and failed
greedy multi-resource scheduling applied to distributed deep learning training clusters. Outcome: worse than baseline. Reason: CPU and memory fragmentation caused severe GPU underutilization under high cluster load
Accelerating deep learning training : a storage perspective · UT Austin
Tried and failed
greedy heuristics for combinatorial scheduling applied to online matching and flow problems. Outcome: worse than baseline. Reason: natural greedy choices fail to achieve optimal competitive ratios, yielding poor worst-case approximation bounds
Designing Networks, Routing Fleets, and Trying to Find Parking · Cornell
Tried and failed
greedy task placement without bandwidth awareness applied to in-network computing resource scheduling. Outcome: worse than baseline. Reason: ignoring bandwidth distribution caused severe performance degradation under high network oversubscription
Data centre resource scheduling for dataflow applications · Imperial
Tried and failed
greedy locality and round-robin placement heuristics applied to batch task scheduling at full capacity. Outcome: worse than baseline. Reason: resource fragmentation and oversubscription under high capacity utilisation
Data centre resource scheduling for dataflow applications · Imperial
Tried and failed
gradient-based greedy resource allocation applied to multicore cache and bandwidth partitioning. Outcome: worse than baseline. Reason: local gradients miss non-convex threshold effects where performance gains require minimum resource thresholds
DYNAMIC RESOURCE ALLOCATION AND ITS APPLICATIONS TO MULTICORE REAL-TIME SYSTEMS · Penn
Tried and failed
classical dispatching heuristic rules for job selection applied to flexible manufacturing system scheduling. Outcome: worse than baseline. Reason: rules ignored load and unload bottlenecks, causing initial machine idle times during startup
Intelligent flexible manufacturing system control · Cranfield
Tried and failed
Static and aggregate heuristic load balancing rules applied to Flexible manufacturing system scheduling. Outcome: unstable. Reason: Ignoring dynamic buffer states caused machine overloading and cascading simulation deadlock.
Simulation and optimisation of a specific flexible manufacturing system. · Cranfield
Tried and failed
weighted shortest remaining processing time rule applied to multi-class job scheduling. Outcome: did not generalise. Reason: Optimality breaks when scaling beyond two distinct priority weight classes.
Scheduling in Healthcare · Cornell
Tried and failed
distance-based priority ordering heuristics applied to makespan minimization scheduling problems. Outcome: did not generalise. Reason: makespan objectives can require visiting farther nodes first, breaking total-variation time ordering assumptions
Min-Time Coverage in Constricted Environments · Georgia Tech
Considered and rejected
Considered and rejected: Rejected greedy bottom-up extraction algorithms from E-graphs in HELM because fabrication manufacturing time is non-additive and requires global operation scheduling.
Programming Language Tools and Techniques for Computational Fabrication · ResearchWorks
Tried and failed
greedy single most likely path scheduling applied to multi-user immersive video streaming. Outcome: worse than baseline. Reason: optimizing only the top trajectory caused high variance and unfair quality of experience across users
Timely information sharing in communication constrained systems · UT Austin
Tried and failed
Greedy incremental resource allocation applied to cycle time constrained vehicle routing. Reason: Non-convex objective returns with respect to integer resource count lead to suboptimal greedy selections.
Tactical design of last mile logistical systems · Georgia Tech
Flawed structural constraints and oversimplified models degrade solution quality
Incorporating invalid symmetry breaking rules or inappropriate structural constraints inadvertently prunes optimal schedules and inflates computational overhead. Simplifying subproblems or imposing overly strict instantaneous bounds degrades overall solution fidelity and network throughput.
Tried and failed
CPU-affinity-aware resource assignment via system calls applied to multithreaded memory allocation. Outcome: worse than baseline. Reason: System call overhead degraded performance compared to simple round-robin assignment
xBGAS-enabled memory centric architecture for high performance computing · Texas Tech
Tried and failed
simple job ordering heuristics applied to DAG task scheduling. Outcome: worse than baseline. Reason: heuristic scheduling produced non-exact schedulability analysis and sub-optimal schedules compared to linear programming formulations
Towards Efficient Autonomous Vehicle Systems: A Multi-Layer Approach · Virginia Tech
Tried and failed
proportional resource allocation based on group sizes applied to network slicing resource management. Outcome: worse than baseline. Reason: asymmetric opportunistic scheduling within slices causes non-linear efficiency differences across groups
Online learning algorithms for wireless scheduling · UT Austin
Tried and failed
hierarchical integer optimization routing and resource allocation applied to sparse satellite constellation network routing. Outcome: worse than baseline. Reason: Off-zenith routing coordination increased path loss, underperforming simple maximum-elevation greedy heuristics in sparse coverage.
Optimizing resource allocation in large communications satellite constellations · MIT
Tried and failed
removing dynamic programming from Lagrangian subproblem applied to flexible manufacturing scheduling optimization. Outcome: worse than baseline. Reason: subproblem simplification significantly degraded the resulting schedule solution quality despite reduced runtimes
Models and algorithms for a flexible manufacturing system · Virginia Tech
Tried and failed
symmetry-breaking constraints in CP-SAT applied to flexible job-shop scheduling. Outcome: worse than baseline. Reason: reification overhead worsened runtime and optimality gap on larger instances
Scheduling in Semiconductor Manufacturing · Georgia Tech
Tried and failed
1:1 family-to-machine assignment symmetry breaking applied to flexible job-shop scheduling. Outcome: worse than baseline. Reason: it pruned improving solutions due to non-identical job release times
Scheduling in Semiconductor Manufacturing · Georgia Tech
Tried and failed
direct randomized rounding of linear programming relaxations applied to ordering and scheduling problems. Reason: fractional coverage across many components leaves items uncovered with high probability due to independence bounds
Randomized Kernel Rounding for Routing, Scheduling, and Machine Learning · Georgia Tech
Tried and failed
gradient-based latency-quality trade-off optimization applied to distributed computation graph scheduling. Outcome: worse than baseline. Reason: analytical fidelity model inaccuracies compounded when relaxing latency constraints and omitting path combination terms
Real-time guarantees in distributed edge computing · UT Austin
Tried and failed
Strict instantaneous interference constraints in resource allocation applied to satellite wireless network scheduling. Outcome: worse than baseline. Reason: Overly strict instantaneous thresholds severely under-utilize capacity compared to time-averaged constraints
Wireless communications in low-earth orbit (LEO) satellite systems · UT Austin
Tried and failed
proportional fair scheduling with multi-connectivity applied to wireless ultra-reliable low-latency communication. Outcome: worse than baseline. Reason: Cross-layer protocol conflicts between MAC bandwidth allocation and packet data convergence protocol path selection
Wireless Network Dimensioning and Provisioning for Ultra-reliable Communication: Modeling and Analysis · Virginia Tech
Tried and failed
multifunctional fleet vehicle allocation applied to shared autonomous vehicle routing. Outcome: worse than baseline. Reason: network connectivity and topological constraints increased fleet size requirements under strict wait-time thresholds
Considered and rejected
Considered and rejected: Standard Vehicle Routing Problem (VRP/VSP) formulation, rejected because constrained LV numbers make fulfilling all rigid time windows infeasible, necessitating processor scheduling formulations with tardiness/lateness objectives
UNDERSEA LOGISTICS: ROUTING OPTIMIZATION IN GRAY ZONE ENVIRONMENTS · Calhoun
Learning-based schedulers suffer from training instability and poor competitive performance
Reinforcement learning and neural network approaches struggle with large action spaces, non-IID sample correlations, and destabilizing reward variances during training. Consequently, these models often fail to converge, exhibit high exploration overhead, or are outperformed by simple heuristic baselines.
Tried and failed
policy gradient with hindsight baseline applied to virtual machine resource allocation. Outcome: worse than baseline. Reason: insufficient variance reduction to compete with model-based hindsight distillation
Adaptivity, Structure, and Objectives in Sequential Decision-Making · Cornell
Tried and failed
proximal policy optimization reinforcement learning applied to sequential dynamic combinatorial resource allocation. Outcome: worse than baseline. Reason: PPO-based policies failed to outperform greedy dynamic heuristics and static integer programming baselines on large instances
Lost to a baseline
RL (Double-DQN, MAC, PPO, and PG with Hindsight Baseline) and standard heuristics (Bin Packing) lost to BestFit baseline on stylized VM allocation (saving negative PMs vs BestFit).
Adaptivity, Structure, and Objectives in Sequential Decision-Making · Cornell
Considered and rejected
Considered and rejected: Rejected direct end-to-end fine-grained GPU allocation in RL agent because joint job-and-GPU selection causes massive action spaces and poor sample efficiency; adopted hierarchical model-selection with heuristic placement instead.
Tried and failed
reinforcement learning for heuristic training applied to dynamic scheduling and task allocation. Outcome: did not converge. Reason: non-IID sample correlation and noisy reward labels during training
Spacecraft Autonomy through Computer Vision and Onboard Planning · MIT
Tried and failed
deep reinforcement learning for dynamic capacity allocation applied to on-demand workforce scheduling. Reason: agent avoided cheaper multi-hour commitments because demand spikes were too short-lived to justify commitment duration
Demand and capacity management for meal delivery systems · Georgia Tech
Tried and failed
average-reward reinforcement learning with dynamic event bounds applied to spatial multi-event scheduling. Outcome: worse than baseline. Reason: large infrequent rewards under high event bounds destabilised reinforcement learning
Towards long-term deployment of general-purpose service robots · UT Austin
Lost to a baseline
On average across 10 runs, the neural network models (e.g., analog Cave = 1785.22 on P1, 4180.7 on P2) did not beat the pseudo-random scheduling heuristic (Cave = 1699.8 on P1, 3905.6 on P2).
A Model for Combinatorial Optimization using Neural Networks and Object-Oriented Programming · TXST Digital Repository
Considered and rejected
Considered and rejected: Rejected Reinforcement Learning (RL) agents for task scheduling due to high exploration overhead, long convergence times in large-scale dynamic environments, and sensitivity to hyperparameter/reward design.
Predictive Worker Resource Characterization at the Extreme Edge · Queens University Institutional Repository
Considered and rejected
Considered and rejected: Rejected machine-learning/deep-reinforcement-learning approaches like FairTS for the core online scheduling algorithms due to inferior computational speed compared to low-complexity heuristics in fog environments.
Scheduling Problems in Next Generation Computing Networks · Queens University Institutional Repository
Tried and failed
primal-dual graph neural network constrained optimization applied to wireless network packet routing and scheduling. Outcome: worse than baseline. Reason: relative efficiency compared to dual descent degrades under heavy traffic congestion loads
GRAPH NEURAL NETWORKS FOR COMMUNICATION IN MULTI-AGENT SYSTEMS · Penn
Mathematical programming formulations suffer extreme computational intractability on large instances
Exact integer and mixed-integer formulations experience exponential growth in variables and constraints as problem scale or network size increases. Solvers timeout, fail to find feasible solutions within operational limits, or prove computationally prohibitive compared to simpler methods.
Considered and rejected
Considered and rejected: Rejected multi-dimensional bin packing algorithms online for dynamic phase reallocation due to high computational overhead; adopted greedy gradient heuristics instead.
DYNAMIC RESOURCE ALLOCATION AND ITS APPLICATIONS TO MULTICORE REAL-TIME SYSTEMS · Penn
Considered and rejected
Considered and rejected: Rejected full combinatorial argmin re-optimization across all client chunk allocations at each completed chunk transfer due to computational complexity, adopting instead an incremental greedy heuristic checking if one additional chunk from the active client decreases the objective.
Tried and failed
robust mixed-integer programming with affine decision rules applied to state-task network scheduling. Outcome: too slow. Reason: the resulting formulations scaled excessively and became computationally intractable compared to deterministic approximations
Formulations, algorithms and software for robust optimisation · Imperial
Tried and failed
exact optimization for multi-machine scheduling applied to multi-work-center job scheduling. Outcome: too slow. Reason: computationally prohibitive to scale beyond three work centers
Assortment Optimization under Customer-driven Substitution · Georgia Tech
Tried and failed
exact integer linear programming schedulability testing applied to multiprocessor real-time task scheduling. Outcome: too slow. Reason: extreme computational complexity makes exact response time analysis impractical beyond 13 tasks
Priority Assignment Algorithms for Real-Time Systems · Virginia Tech
Tried and failed
exact integer programming on time-space network applied to large-scale sort network scheduling. Outcome: too slow. Reason: computational complexity caused solver timeouts without finding feasible solutions on large problem instances
Sort Planning for Express Parcel Delivery Systems · Georgia Tech
Considered and rejected
Considered and rejected: Rejected the flow shop scheduling constructive heuristics of Framinan & Leisten and Laha & Sarin due to their excessive computational runtime complexity of O(m · n^4)
Entwicklung eines hybriden Optimierungsverfahrens für dynamische multikriterielle Produktionsglättung am Beispiel der Lebensmittelindustrie · DSpace-CRIS at TU Wien
Tried and failed
direct mixed-integer linear programming solver applied to large-scale transit resource allocation. Outcome: too slow. Reason: computational complexity prevented finding feasible solutions within the time limit on large instances
Fair and Risk-Averse Resource Allocation in Transportation Systems under Uncertainties · Virginia Tech
Tried and failed
mixed integer linear programming space-time network formulation applied to real-time routing and scheduling on large networks. Outcome: too slow. Reason: curse of dimensionality causes exponential growth of variables and constraints, making real-time solving computationally intractable
Integrated and joint optimisation of runway-taxiway-apron operations on airport surface · Imperial
Considered and rejected
Considered and rejected: Rejected simultaneously optimizing truck routing, scheduling, and cumulative hub capacity in a single monolithic MIP/CP due to lack of scalability at national scale.
Autonomous Transfer Hub Networks for Self-Driving Trucks · Georgia Tech
Static and offline policies fail under dynamic runtime variations and uncertainty
Allocations based on fixed thresholds, offline topologies, or deterministic linear programs fail to adapt to time-varying arrivals, non-stationary demand, and dynamic environment shifts. Ignoring runtime state transitions and uncertainty leads to frequent constraint violations, underutilized resources, and missed deadlines.
Tried and failed
static proportional heuristic resource partitioning applied to heterogeneous dynamic traffic multiplexing. Outcome: worse than baseline. Reason: inflexible allocation heuristics fail to adapt to mixed traffic demands compared to dynamic reinforcement learning
Real-Time Resource Optimization for Wireless Networks · Virginia Tech
Tried and failed
pre-allocating resources based on returning entity predictions applied to online bipartite matching. Outcome: worse than baseline. Reason: locking allocations based on predicted returns reduced flexibility compared to greedy load balancing
Tried and failed
Unit-utility renewal analytic heuristic for lookahead scheduling applied to dynamic task scheduling. Outcome: worse than baseline. Reason: Unit-utility assumption broke down under Pareto-distributed task utility, causing severe utility loss
Spacecraft Autonomy through Computer Vision and Onboard Planning · MIT
Tried and failed
static rule-based scheduling policies applied to multi-user wireless network scheduling. Outcome: did not generalise. Reason: fixed parameters failed to adapt across diverse and dynamic channel environment scenarios
Online learning for scheduling in wireless networks · UT Austin
Tried and failed
static policy execution from deterministic linear programs applied to dynamic resource allocation with nonstationary demand. Outcome: worse than baseline. Reason: static schedules fail to adapt to time-varying arrival distributions without price reservation thresholds
Data Driven Operations: From Algorithm Development to Experimental Design · MIT
Tried and failed
static threshold reservation policies applied to multi-class resource allocation. Outcome: worse than baseline. Reason: static rules cannot adapt to time-varying state transitions, making them provably suboptimal versus dynamic state-dependent policies
Essays in Healthcare Operations and Management · Georgia Tech
Tried and failed
deterministic optimal scheduling under uncertainty applied to coupled multi-energy network dispatch. Outcome: did not generalise. Reason: ignoring stochastic uncertainty led to frequent voltage and temperature constraint violations in realization scenarios
Co-optimisation of electrical and district heating networks · Imperial
Tried and failed
replacing dynamic scheduling with static empirical distribution applied to distributed task scheduling. Outcome: worse than baseline. Reason: online dynamic state awareness cannot be substituted by long-term empirical sampling distributions
Principled Approaches for Latency Reduction in Networking Systems · MIT
Considered and rejected
Considered and rejected: Offline static topology scheduling was rejected because it fails to adapt dynamically to runtime input fluctuations and leaves unassigned worker slots underutilized
QoS-aware Resource-utilisation Self-adaptive (QRS) Framework for Distributed Data Stream Management Systems · De Montfort Open Research Archive (DORA)
Tried and failed
real-time task scheduling with linear optimization applied to vehicular edge computing task offloading. Reason: speed and motion uncertainties cause deadline misses in realistic dynamic driving conditions
Predictable Connected Traffic Infrastructure · Virginia Tech
Iterative optimization algorithms and metaheuristics fail to converge reliably
Continuous solvers and metaheuristic search methods frequently fail when applied to complex multi-period or non-linear scheduling problems. Algorithms oscillate without reaching stationary profiles, converge prematurely without diversity, or fail to find optimal values due to objective discontinuities and inappropriate penalty parameters.
Tried and failed
metaheuristic optimization algorithms applied to dynamic sequential tactical resource allocation. Outcome: did not generalise. Reason: they lack dynamic feedback loops needed for co-evolving, semi-adversarial simulation environments
A Simulation-Based Methodology for Sequential Decision-Making in Dynamic and Uncertain Tactical Environments · Georgia Tech
Tried and failed
genetic algorithm for mixed-integer nonlinear programming applied to large-scale parallel machine scheduling. Outcome: infeasible cost. Reason: High problem complexity caused constraint violations and intractable computational scaling for long time horizons.
Probabilistic Approaches to Enhance Safety and Energy Management of Energy Transition Components and Systems · Texas Tech
Tried and failed
genetic algorithms with discrete event simulation applied to constrained campaign scheduling optimization. Reason: Lacked optimality guarantees, suffered high computational cost, and required ad-hoc penalty functions for complex logical constraints.
Development and Acquisition Modeling for Space Campaign Architecting · Georgia Tech
Tried and failed
simple genetic algorithm without diversity filtering applied to power unit commitment scheduling. Reason: converged prematurely to a single optimum, failing to identify diverse alternative operational solutions
Genetic algorithm unit commitment program · Iowa State
Tried and failed
Sequential optimization without constraint screening applied to multiperiod network power flow scheduling. Outcome: did not converge. Reason: Iterative updates oscillated across iterations without reaching a stationary loss profile
Generalized Energy Resource Scheduling for Distribution Grid Operations Planning · Georgia Tech
Tried and failed
interior-point method with switching objective functions applied to optimal power flow scheduling. Outcome: did not converge. Reason: discontinuities from switching objective formulations caused convergence failure across varying active objective sets
Smart electric vehicle charging strategy in direct current microgrid · Imperial
Tried and failed
dual stochastic dual dynamic programming with exact penalization applied to hydrothermal power generation scheduling. Outcome: did not converge. Reason: fixed penalty parameter was too small to reach the optimal value
Risk neutral and risk averse stochastic optimization · Georgia Tech
Lost to a baseline
Standard PSO was outperformed in solution accuracy and convergence rate by the proposed modified PSO (with linear decreasing inertia weight and periodic mutation) across 30 runs on the multi-microgrid scheduling problem.
Evolutionary Algorithms for Resource Allocation in Smart Grid · De Montfort Open Research Archive (DORA)
Left open by the authors
Problems the authors named and did not get to.
Left open
Rigorously prove that informative scheduling policies stochastically outperform non-informative counterparts across general queueing settings. Blocker: None
Information Freshness Optimization in Real-time Network Applications · Virginia Tech
Left open
Use PerformanceLENS latency predictions to automate microservice resource allocation in cloud-native environments. Blocker: No specific resource allocation algorithm, policy, target, or evaluation framework is defined
Left open
Optimize Juice incremental snapshots using contiguous allocation heuristics to balance heap versus dirty heap tracking costs. Blocker: None
Simpler & Safer Programming Models for Web Infrastructure & Applications · Harvard
Left open
Derive optimal analytical allocation rules for the computational budget across the planning horizon in sequential sample average approximation. Blocker: None
ONLINE LEARNING OF MARKOVIAN SYSTEMS WITH CENSORED POISSON ARRIVALS · Calhoun
Left open
Optimize dynamic computation and upload budget allocation across heterogeneous workers based on delay statistics for SBP and MM-GASP codes. Blocker: None
Efficient, secure and private distributed computation and machine learning · Imperial
Left open
Determine best-fit resource allocation metrics for in situ monitoring platforms across major compute clusters. Blocker: The unfinished work lacks a specific methodology, target metrics, or criteria beyond a general research direction.
Scalable Observation, Analysis, and Tuning for Parallel Portability in HPC · Scholars' Bank
Left open
Add resource utilization rate calculation to evaluate effects of supply adjustments in the scheduling heuristic. Blocker: None
Left open
Adapt discretized metaheuristic algorithms to solve dynamic nurse scheduling problems with unexpected staff absences and demand variations. Blocker: None
Discretization of nature-inspired techniques for combinatorial problems · oURspace
Left open
Adapt HetGPO graph attention scheduling architecture to optimize hospital patient admission scheduling problems. Blocker: None
Learning Dynamic Priority Scheduling Policies with Graph Attention Networks · Georgia Tech
Left open
Investigate compute-efficient methods to reduce training time for reinforcement learning agents in reconfigurable manufacturing system scheduling. Blocker: No specific algorithmic approach or concrete efficiency target is specified
A complete reinforcement learning based framework for reconfigurable manufacturing system scheduling · Cranfield
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.