Chapter Four · failure evidence
What Decentralized & Distributed Control got wrong, from 97 dissertations
The records examine trade-offs between centralized and distributed control architectures across multi-agent robotics, power grids, and networked compute systems. While centralized schemes encounter computational bottlenecks and single points of failure, decentralized approaches often suffer from communication latency, non-stationarity, local deadlocks, and sub-optimal performance. These records come from PhD theses at 28 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.
Communication latency and message passing overhead degrade distributed and parallel control performance
Distributed and parallelized schemes frequently suffer from network delays, packet loss, and high message passing overhead that negate parallelization speedups or induce dynamic instability. In several multi-agent and power system settings, synchronization and consensus fail when communication latencies exceed reporting thresholds or when link bandwidth is constrained.
Tried and failed
distributed parameter tracking for dynamic coordination applied to distributed grid frequency control. Outcome: unstable. Reason: distributed control topology exhibits higher dynamic sensitivity to communication latency despite smaller absolute delay magnitudes
Frequency Control of Virtual Power Plants · Research Repository UCD
Tried and failed
uplinking linearized model matrices during control loop applied to distributed ground-in-the-loop control systems. Outcome: worse than baseline. Reason: uplinking intermediate linearizations was consistently suboptimal across latency and communication bandwidth trade-off trials
Embedded Computing for Wavefront Control on Future Space Telescopes · MIT
Tried and failed
dynamic load balancing on small core counts applied to distributed parallel structure learning. Outcome: worse than baseline. Reason: communication and redistribution overhead exceeded gains from mitigating load imbalance at small scales
Parallel Algorithms and Generalized Frameworks for Learning Large-Scale Bayesian Networks · Georgia Tech
Lost to a baseline
In dense communication graph settings (avg 25 neighbors/node), centralized graph learning methods match or slightly outperform the distributed method in communication costs.
Quantization for Distributed Processing and Learning of Structured Data · EPFL
Lost to a baseline
Centralized MPC was 2.8x faster (2.0s vs 5.6s mean time) and had lower dispatch RMSE (1.6% vs 3.18%) than distributed ADMM MPC
Methods for Grid-aware Operation and Planning of Active Distribution Networks · EPFL
Considered and rejected
Considered and rejected: Rejected distributed scheduling algorithms in favor of centralized GPU base-station scheduler to reduce unpredictable communication overhead
Coping Uncertainty in Wireless Network Optimization · Virginia Tech
Considered and rejected
Considered and rejected: Decentralized peer-to-peer cluster control (as in HomeEdge) rejected in favor of centralized cluster heads due to vulnerability to malicious code execution and O(N^2) benchmark message overhead.
Community-Oriented Edge Computing Platform · Queens University Institutional Repository
Tried and failed
decentralized model predictive control with asynchronous communication applied to multi-agent collision avoidance. Outcome: unstable. Reason: asynchronous communication and optimization latency failed to prevent collisions at high agent densities
Microscopic analysis of many optimizing air vehicles using high-performance computing · Georgia Tech
Lost to a baseline
Onboard distributed MPC at dt = 50 ms had a higher constraint violation and collision rate (3 collisions) compared to dt = 150 ms (1 collision) or offboard execution (0 collisions) due to fewer real-time feasible ADMM iterations and wireless communication delays
Numerical Methods for Optimization and Control: From Verification to GPU-Acceleration · EPFL
Considered and rejected
Considered and rejected: Decentralized interactive active labeling across all robots was rejected in Chapter 4 due to excessive communication overhead, leading to the decoupled two-stage DUAL framework
Data collection for robust machine learning in multi-agent systems · UT Austin
Considered and rejected
Considered and rejected: Rejected DDS/multicast-based distributed discovery in SM2 due to performance degradation and unreliability over lossy multi-robot wireless links, choosing centralized Manager registration instead.
Toward a Cloud Continuum for Efficient 6G Infrastructures · IRIS - POLITO - prod
Tried and failed
multithreading and distributed computing for parallel sensitivity evaluation applied to nonlinear model predictive control sensitivities. Outcome: too slow. Reason: thread spawning and communication overhead outweighed parallel computational speedups
Real-Time Nonlinear Model Predictive Control for Fast Mechatronic Systems · EPFL
Tried and failed
cyber-physical stability boundary mapping under sampled latency applied to microgrid distributed secondary control. Outcome: unstable. Reason: sampling reporting period exceeding critical threshold causes instability regardless of minimal communication delay
Microgrid as a Cyber-Physical System: Dynamics and Control · Virginia Tech
Tried and failed
distributed parallel computing on low-power architectures applied to distributed fast Fourier transform workloads. Outcome: too slow. Reason: communication overhead and low per-core throughput prevented effective parallel speedup and degraded efficiency
The Performance Comparisons of Single Board Computer Clusters · Texas Tech
Tried and failed
uncoded multi-message transmission scheduling applied to distributed parameter server communication. Outcome: worse than baseline. Reason: high network latency and delay probability caused excessive communication overhead and parameter server congestion
Moving intelligence to edge: distributed content storage and computation · Imperial
Tried and failed
distributed decision tree induction using MapReduce applied to tabular classification on multi-node clusters. Outcome: too slow. Reason: communication overhead and coordination bottlenecks offset parallelization gains on small clusters
Scaling MapReduce-based big data analytics algorithms with AI inspired controls · Texas Tech
Tried and failed
distributed MPI parallelization with fixed problem size applied to hyperbolic PDE network simulation. Outcome: too slow. Reason: inter-processor communication overhead dominates computational work at high processor counts under strong scaling
Advancing water distribution systems modeling and management through high-performance computing · UT Austin
Tried and failed
fine-grained dynamic distributed matrix operation scheduling applied to massive MIMO baseband processing. Outcome: too slow. Reason: Inter-server network communication overhead outweighed computation gains, which single modern CPU cores handled adequately.
Managing Virtualized Network Functions in the Cloud · Harvard
Tried and failed
distributed memory parallelization via node partitioning applied to finite element method simulations. Outcome: too slow. Reason: communication overhead exceeded local computation as processor count increased for small workloads
Towards automatic parallelization of finite element method · Iowa State
Considered and rejected
Considered and rejected: Rejected distributed secondary control relying on digital communication links (DCL) due to communication latency, traffic, cyber-attack risks, and single points of failure in flight-critical MEA systems.
Control of electrical systems using artificial intelligence for more electric aircraft applications · University of Nottingham Repository
Tried and failed
distributed consensus algorithms applied to multi-agent coordination under packet loss. Outcome: did not converge. Reason: communication packet loss breaks convergence to the global balance constraint
Tried and failed
corralling bandits for model selection applied to decentralized robust multi-agent bandits. Outcome: did not converge. Reason: communication delays and absence of shared randomness prevent coordination under unknown contamination
Private and Provably Efficient Federated Decision-Making · MIT
Considered and rejected
Considered and rejected: Decided against extending epsilon-BAI to distributed multi-agent service placement because coordinating consensus on local epsilon-optimality and arm elimination across heterogeneous agents significantly increases communication complexity.
Bandit Learning for Resource Allocation and Performance Optimization in Edge Networks · Publikationssystem UB Tuebingen
Lost to a baseline
Centralized learning achieved lower latency (11.92 hours vs 36.58 hours for FedProx Sequential) due to avoiding coordination and proximal term overhead
Privacy-Preserving Federated Learning for Secure and Scalable Digital Pathology · DSpace at SUNY Buffalo
Considered and rejected
Considered and rejected: Rejected hybrid and decentralized/distributed testing architectures in favor of a purely centralized test architecture to avoid additional message-passing overhead and coordination complexity on resource-constrained IoT devices.
Regression Testing of Distributed Real-Time Embedded Systems in the Context of Model-Driven Development · Queens University Institutional Repository
Centralized control architectures present single points of failure and physical implementation vulnerabilities
Centralized controllers and orchestrators are frequently rejected because a single failure or cyber targeting event can compromise the entire operational network. Designers also reject centralized setups due to excessive physical cabling weight, long-distance communication dependencies, and the loss of local authority.
Tried and failed
centralized communication network architecture applied to multi-agent coordination systems. Outcome: worse than baseline. Reason: high centralization reduced group cohesiveness and overall system performance compared to decentralized networks
Measuring and Enhancing the Resilience of Interdependent Power Systems, Emergency Services, and Social Communities · Virginia Tech
Tried and failed
hierarchical centralized command structures applied to asymmetric multi-agent coalition coordination. Outcome: worse than baseline. Reason: centralized hierarchical control degraded operational performance rather than improving coordination in asymmetric partnerships
Considered and rejected
Considered and rejected: AC-based centralized controller nanogrid architecture: despite scoring highest in the initial Pugh matrix, it was rejected by base managers due to survivability concerns with centralized control and preference for DC/decentralized architectures
INVESTIGATION OF NANOGRIDS FOR IMPROVED NAVY INSTALLATION ENERGY RESILIENCE · Calhoun
Considered and rejected
Considered and rejected: Centralized global controller rejected because it demands long-distance communication between converters and a single point of failure compromises the entire grid.
Optimal H2 control design of active front-end integrating grid model identification · University of Nottingham Repository
Considered and rejected
Considered and rejected: Wide area centralized control rejected in favor of distributed local control due to prohibitive communication infrastructure costs and cyber-vulnerability.
Power System Stability Improvement with Decommissioned Synchronous Machine Using Koopman Operator Based Model Predictive Control · Virginia Tech
Considered and rejected
Considered and rejected: Rejected centralized NCS control due to vulnerability of single controller failure/targeting in adversarial environments, adopting distributed submodularity.
RED CELL ANALYSIS FOR MOBILE NETWORKED CONTROL SYSTEMS · Calhoun
Considered and rejected
Considered and rejected: Centralized microgrid power management relying on 'power available' calculations was rejected due to single-point-of-failure vulnerability
Variable Voltage-based Power Management of DC Microgrids : With Focus on Shipboard Power Systems · Osuva
Considered and rejected
Considered and rejected: Rejected centralized submodularity method for the Docker container implementation because it relies on a single controller rather than true distributed execution.
INCREASING UTILITY AND FIDELITY OF A UXV NETWORKED CONTROL SYSTEM SIMULATION VIA PYTHON ADAPTATION · Calhoun
Considered and rejected
Considered and rejected: Rejected centralized MINLP optimization for multi-robot tracking because it is NP-hard and prone to single-point failures, opting for two-stage decentralized coordination and control decomposition.
HOLISTIC SCENE PERCEPTION FOR COLLABORATIVE HUMAN-ROBOT TEAMS · Cornell
Considered and rejected
Considered and rejected: Rejected centralized orchestrator architectures in favor of decentralized BPMN engines per robot to prevent a single point of failure.
Process-driven Development and Analysis of Multi-Robot Systems · IRIS - UNICAM - prod
Considered and rejected
Considered and rejected: Rejected centralized MPC architectures due to single-point failure vulnerability, high computational demands, and impractical continuous communication needs in unpredictable environments
Considered and rejected
Considered and rejected: Centralized MPC architecture rejected due to communication single-point-of-failure and latency.
Towards Safe Autonomous Multirotor Landing on Uncrewed Surface Vessels in Waves using Model Predictive Control · Queens University Institutional Repository
Considered and rejected
Considered and rejected: Rejected centralized LQR and centralized MPC control architectures due to single-point-of-failure vulnerabilities, communication bandwidth bottlenecks, and excessive cabling weight in MEA.
The optimal control of power electronic embedded networks in More Electric Aircraft · University of Nottingham Repository
Tried and failed
centralized administrative restructuring of decentralized entities applied to autonomous regional transport infrastructure. Reason: local authorities resisted loss of autonomy, leading to institutional collapse and dissolution within a year
Considered and rejected
Considered and rejected: Centralized global localization (MCS/GNSS-dependent control) was rejected in favor of onboard relative sensing to enable GPS-denied autonomous operation.
Distributed Predictive Formation Control of Autonomous Rotary-Wing Micro Aerial Vehicles · EPFL
Considered and rejected
Considered and rejected: Rejected centralized information sources in multi-sensor active planning, adopting instead a distributed geometric-average Kalman consensus filter.
Distributed, Scalable And Resilient Information Acquisition For Multi-Robot Teams · Penn
Local constraint conflicts, deadlocks, and information inconsistency undermine decentralized coordination
Distributed and decentralized coordination algorithms frequently fail when agents encounter conflicting safety constraints, spatial deadlocks, or inconsistent local beliefs. Uncoordinated local updates and unconstrained boundary conditions can lead to empty admissible control sets, goal oscillations, or system collapse.
Tried and failed
decentralized multi-agent planning under low communication stochasticity applied to multi-robot inspection coordination. Outcome: worse than baseline. Reason: Low communication dropouts caused multi-agent deadlocks, whereas higher stochasticity helped break symmetrical decision deadlocks.
Sequential decision-making and active sensing for autonomous outer space operations · UT Austin
Tried and failed
purely centralized or decentralized coordination architectures applied to autonomous vehicle collision avoidance. Reason: centralized creates controller bottlenecks while decentralized suffers high coordination latency and collision risk under dense traffic
A Systems-Theoretic Framework For Safety-Driven Development of System Architectures · MIT
Tried and failed
decentralized pairwise control barrier functions applied to multi-agent collision avoidance. Reason: conflicting pairwise safety constraints produce an empty intersection of admissible controls
Barrier Functions and Model Free Safety With Applications to Fixed Wing Collision Avoidance · Georgia Tech
Tried and failed
artificial zero-mean dither perturbation in decentralized control applied to modular parallel power converters. Outcome: did not converge. Reason: dither amplitude was smaller than measurement discrepancies, preventing consensus on error signs across decentralized units
Tried and failed
Distributed potential field control using interoceptive sensing applied to tethered multi-agent formation control. Outcome: unstable. Reason: Conflicting angle changes with unconstrained boundary agents caused inward collapse and crossed connections or drift
An Examination of How Tethers Can Support Collaborative Behaviors in Multi-Robot Systems · Cornell
Tried and failed
distributed naive averaging applied to multi-agent data fusion. Outcome: unstable. Reason: incapacity to handle common information consistently across nodes
Bayesian Learning: Paving the Way to Trustworthy Robots · Georgia Tech
Tried and failed
distributed belief propagation without occupancy weighting applied to multi-agent goal assignment and coordination. Outcome: unstable. Reason: agents oscillate between goal choices when neighbors drop out of communication range
Distributed coordination and perception for multi-robot systems · Imperial
Lost to a baseline
Distributed fictitious play failed to converge within 500 steps in 18 of 50 runs on geometric networks and 5 of 50 runs on small-world networks.
Bayesian Network Games · Penn
Considered and rejected
Considered and rejected: Rejected purely relative state feedback in consensus networks due to unbounded state variance growth under distributed Wiener noise.
Analysis, Control, and Optimization of Networked Dynamical Systems under Uncertainty · JScholarship
Tried and failed
rule-based car-following and lane-changing models applied to cooperative multi-agent vehicle overtaking maneuvers. Outcome: did not generalise. Reason: heuristic decentralized controllers lack mechanisms for multi-agent coordinated yielding and cooperation
Semi-Cooperative Planning in Mixed Human-Autonomous Environments · MIT
Tried and failed
cooperative routing with decentralized consensus applied to mixed autonomous and human traffic. Outcome: did not generalise. Reason: uncontrolled non-cooperative human agents invalidate static multi-agent coordination assumptions
Incentive Mechanisms for Collaborative Routing Factoring Individual Heterogeneities · Georgia Tech
Tried and failed
distributed decision-making without explicit escalation boundaries applied to organizational leadership structure. Outcome: unstable. Reason: Unclear authority boundaries caused role anxiety, reverting the system to centralized control.
Tried and failed
decentralized uncoordinated tracking control applied to multi-target multi-agent tracking. Outcome: did not generalise. Reason: failed to adapt to target crossings, leaving distant targets unattended
HOLISTIC SCENE PERCEPTION FOR COLLABORATIVE HUMAN-ROBOT TEAMS · Cornell
Lost to a baseline
Conviction Voting (continuous decentralized preference signaling) was abandoned in favor of traditional centralized decision-making due to lack of coordination and political coherence.
Emergence of the Crypto Commons. Navigating Socio-Technical Affordances and Ideological Tensions on the Blockchain. · IRIS - SNS - prod
Considered and rejected
Considered and rejected: Rejected track-to-track fusion architecture to prevent rumor propagation feedback loops across vehicles, opting instead for centralized detection-level fusion at the ego vehicle.
Enhancing Perception Systems using V2V Sensor Fusion · Virginia Tech
Centralized optimization encounters computational bottlenecks and exponential scaling limits
Centralized controllers encounter severe computational bottlenecks and state-action explosions when handling large agent fleets or dense real-time optimization problems. These computational and service-request delays prevent real-time trajectory verification, motion planning, and optimal resource dispatch.
Tried and failed
centralized continuous reinforcement learning applied to multi-agent spatial resource allocation. Outcome: worse than baseline. Reason: centralized control scaled poorly compared to multi-agent distributed strategies in large action spaces
Modelling urban street configurations for Autonomous Vehicle flows · Imperial
Lost to a baseline
Decentralized symbolic plan execution achieved lower per-period computation time (0.021 s) compared to centralized dynamic assignment (0.033 s)
Considered and rejected
Considered and rejected: Rejected centralized smart grid SCADA architectures because communication latency and processing bottlenecks prevent real-time DER optimization
Considered and rejected
Considered and rejected: Centralized DRF rejected for distributed multi-controller SDN due to poor scalability across independent administrative domains.
A Game-Theoretic Approach to Strategic Resource Allocation Mechanisms in Edge and Fog Computing · De Montfort Open Research Archive (DORA)
Considered and rejected
Considered and rejected: Centralized load-shedding control architecture rejected due to communication latency, complex processor logic, and computational bottlenecks in dynamic environments.
A new automatic under-frequency load shedding scheme New automatic under frequency load shedding scheme · Iowa State
Tried and failed
centralised coordination of autonomous mobile agents applied to multi-robot fleet management. Outcome: too slow. Reason: central controller encountered computational bottlenecks while underutilising distributed edge compute capacity
Mobile robot automation in warehouses. · Cranfield
Considered and rejected
Considered and rejected: Rejected centralized optimal motion planning for formations of n > 2 robots due to prohibitive computational complexity, adopting decentralized leader-follower control graphs instead.
Motion Planning and Control of Cooperative Robotic Systems · Penn
Considered and rejected
Considered and rejected: Rejected centralized optimal control with full-space joint optimization due to exponential scaling failure with multi-agent traffic.
Applying FastMDP to complex aerospace-related problems · Iowa State
Considered and rejected
Considered and rejected: Rejected centralized intersection management architecture due to exponential computation time scaling with vehicle count and single-point-of-failure vulnerabilities, selecting distributed/decentralized control instead.
Considered and rejected
Considered and rejected: Rejected centralized joint HPP-MDP multi-robot planning due to exponential state-action explosion, replacing it with a decentralized variance-rewarding bidding mechanism
Tried and failed
centralized data-driven predictive control applied to multi-agent trajectory planning. Outcome: too slow. Reason: computational bottleneck from solving large dense optimization matrices in real time
Hierarchical Control of Constrained Multi-Agent Legged Locomotion: A Data-Driven Approach · Virginia Tech
Considered and rejected
Considered and rejected: Rejected treating vehicle and parcel routes as fully centralized dynamic optimization models due to intractable real-time computational complexity; adopted decentralized agent-based routing over multi-tier containment layers instead
Hyperconnected Parcel Logistics: Planning and Assessment · Georgia Tech
Considered and rejected
Considered and rejected: Rejected a centralized trajectory-path verification ATM service due to service-request bottlenecks and latency in favor of a decentralized architecture where the service only distributes building database subsets to vehicles.
Microscopic analysis of many optimizing air vehicles using high-performance computing · Georgia Tech
Considered and rejected
Considered and rejected: Rejected centralized cloud architectures for O-DRL deployment in favour of a decentralized edge computing architecture with local intersection controllers and edge servers.
Urban Traffic Management Using Deep Reinforcement Learning in Connected Vehicle Environments · Research Repository UCD
Decentralized and decoupled control yields sub-optimal performance and higher operating losses
Operating controllers in localized, decoupled, or uncoordinated modes neglects system-wide interdependencies and global constraints. As a result, decentralized methods suffer from increased power losses, higher life-cycle costs, sub-optimal routing, and degraded service quality compared to centralized baselines.
Tried and failed
direct transmission grid integration without frequency control applied to distributed energy resources dynamic response. Outcome: worse than baseline. Reason: deteriorates system dynamic response worse than aggregating resources at the distribution level
Frequency Control of Virtual Power Plants · Research Repository UCD
Tried and failed
decentralized droop control without secondary coordination applied to heterogeneous multi-source power sharing. Reason: unequal power sharing due to differing dynamics and lack of global secondary voltage/frequency regulation
Resilience and Cybersecurity for Distribution Systems with Distributed Energy Resources · Virginia Tech
Tried and failed
decoupled subproblem optimization applied to joint resource allocation in distributed networks. Outcome: worse than baseline. Reason: optimizing coupled parameters in isolation ignores inter-variable dependencies, yielding degraded throughput and higher power consumption
Optimizing resource allocation in large communications satellite constellations · MIT
Lost to a baseline
On the 33-bus power network, both SMFPPO and SQDDPG achieved higher power loss than centralized OPF (though lower than Droop Control).
Lost to a baseline
Decentralized ES-tuned droop yielded 86.18 kWh network loss and 0.0089 V voltage deviation, losing to idealized centralized OPF which achieved 34.32 kWh loss and 0.0037 V deviation on IEEE 123-node system.
Lost to a baseline
Centralized architectures yielded the lowest 25-year Life Cycle Cost ($17.74M for 4x 2500 kW vs. $17.99M–$20.56M for decentralized clusters), beating decentralized options on raw cost alone.
TRADEOFF ANALYSIS OF BACKUP POWER GENERATION SOLUTIONS FOR MILITARY BASES · Calhoun
Lost to a baseline
Distributed chordal-decomposed OTSP produced a sub-optimal graph with total edge weight cost 4.0979 compared to 3.7644 achieved by centralized optimal MISDP
Distributed, Stable Topology Control of Multi-Robot Systems with Asymmetric Interactions · Virginia Tech
Considered and rejected
Considered and rejected: Operating inverters in localized control mode (e.g., local unity power factor mode) post-attack mitigation was rejected as a long-term solution because localized control is sub-optimal compared to centralized optimal power flow.
AI-based Detection Against Cyberattacks in Cyber-Physical Distribution Systems · Virginia Tech
Lost to a baseline
Decentralized auction-based sub-swarm assignment resulted in larger average travel distance per robot compared to optimal centralized integer programming across swarm sizes 10 to 70
Lost to a baseline
Condition m2 + m4 + m6 < 2 for balanced DHV is weaker than Dai et al. (1999)'s global stability condition m2 + m4*m6 < 1 due to restricting to decentralized queue-ratio policies vs centralized non-idling policies
Policy robustness : robust stability of multiclass queueing networks · UT Austin
Lost to a baseline
The distributed greedy association approach incurs a minimum user throughput penalty of up to 46% for SDFOV, 16% for DFOV, and 57% for FFOV compared to the centralized coordinated optimization.
Coexistence of directional and non-directional technologies in 6G wireless dense networks · OpenBU
Lost to a baseline
Distributed trust architectures (Scheme-2 and Scheme-3) were beaten by centralized architecture (Scheme-1) in maintaining low attacker trust under medium traffic density.
Lost to a baseline
In cell-free pilot allocation with a high user-to-pilot ratio, the RSA-based scheme was marginally beaten by the centralized max-min distance-based scheme due to reduced pilot assignment probability.
Stochastic Geometry Perspective of Massive MIMO Systems · Virginia Tech
Lost to a baseline
Tercio-based centralized baseline achieved fewer late tasks on heterogeneous teams across n=4, 5, 6 robots compared to the proposed ACS reputation approach.
Enabling collaborative autonomy: from human-robot teams to multi-agent systems · OpenBU
Environment non-stationarity and lack of coordination destabilize decentralized learning
Multi-agent reinforcement learning and independent policy optimization frequently fail or become unstable when agents update their policies concurrently without shared coordination. These approaches struggle with partial observability, conflicting world models, and performance collapse across heterogeneous agent environments.
Tried and failed
transferring single-agent curriculum policy to multi-agent settings applied to multi-agent robotic control. Outcome: did not generalise. Reason: lack of inter-agent coordination mechanisms during decentralized execution
Efficient Reinforcement Learning for Control · Virginia Tech
Tried and failed
Multi-Agent Deep Deterministic Policy Gradient applied to multi-robot decentralized sequential decision-making. Outcome: worse than baseline. Reason: struggled with partial observability and coordination under uncertainty despite extensive hyperparameter tuning
Tried and failed
independent world models in decentralized agents applied to multi-agent coordination. Outcome: did not converge. Reason: agents develop conflicting, unsynchronized state representations without explicit coordination structures
Artificial General Intelligence (AGI)-Native Wireless Systems: Digital Twins and World Models for Beyond 6G Networks · Virginia Tech
Considered and rejected
Considered and rejected: Single-agent and decentralized multi-agent RL formulations rejected because they struggle to achieve fleet-wide coordination and global efficiency
Real-Time Idle Vehicle Relocation and Dynamic Pricing in Centralized Ride-Hailing Systems · Georgia Tech
Tried and failed
decentralized training of multi-agent reinforcement learning applied to multi-intersection traffic signal control. Outcome: unstable. Reason: environment non-stationarity caused by independent agents updating policies concurrently without coordination during training
Adaptive Transit Signal Priority based on Reinforcement Learning, Connected Vehicles and Software in the Loop Simulation · Georgia Tech
Considered and rejected
Considered and rejected: Rejected relying purely on individual time-delay (selfish) penalty rewards without global reward mixing and exit bonuses in distributed mixed-autonomy driving.
Communication and generalization in multi-agent learning · UT Austin
Tried and failed
independent Q-learning in multi-agent environments applied to decentralized multi-agent control. Outcome: unstable. Reason: environment non-stationarity caused by independent simultaneous policy updates across interacting agents
Learning-based decision making for safe and scalable autonomous separation assurance · Iowa State
Tried and failed
centralized reinforcement learning with distributed execution applied to multi-user wireless beamforming coordination. Outcome: unstable. Reason: difficulty learning shared representations across highly diverse heterogeneous agent environments causing performance collapse
Optimization and reinforcement learning for emerging multi-antenna wireless networks · Imperial
Considered and rejected
Considered and rejected: Centralized training with decentralized execution was rejected to ensure independent learning without shared action coordination during learning.
Decentralized Learning in Stochastic and Mean-Field Games · Queens University Institutional Repository
Left open by the authors
Problems the authors named and did not get to.
Left open
Model and evaluate cyber latency stability regions for a microgrid including multiple synchronous generators, grid-forming inverters, and grid-following inverters under centralized control. Blocker: None
Microgrid as a Cyber-Physical System: Dynamics and Control · Virginia Tech
Left open
Implement distributed optimization schemes such as consensus+innovations for power grid multi-agent resilience evaluation on IEEE standard test systems. Blocker: None
Measuring and enhancing resilience of an increasingly multi-agent power grid · UT Austin
Left open
Develop decentralized or market-based operational frameworks such as peer-to-peer trading or distributed control for coupled electricity and district heating networks. Blocker: The unfinished work describes broad research directions rather than a specific formulation or target mechanism
Co-optimisation of electrical and district heating networks · Imperial
Left open
Study cloud computing architectures and standardize control partitioning for multi-processor real-time distributed power converter control systems. Blocker: Requires multi-processor power electronics hardware and lacks a specific methodology or concrete research target.
Synchronized Communication Network for Real-Time Distributed Control Systems in Modular Power Converters · Virginia Tech
Left open
Extend grid-forming inverter control strategies to coordinate solar, wind, and EV distributed energy resources collectively across large-scale power grid simulations. Blocker: None
Grid Forming Inverter Control for Inverter Based Resources (IBR) Dominated Power Grid · Texas Tech
Left open
Incorporate distribution network power flow constraints and network losses into the consensus-based distributed energy resource coordination algorithm. Blocker: None
Left open
Extend the unified control strategy to distributed generation converters like wind power and test experimentally on physical hardware. Blocker: Requires a physical hardware microgrid/SST laboratory apparatus and DG converters for experimental testing.
Unified Control Strategy for Microgrid Solid-State Transformers · YorkSpace
Left open
Apply distributed optimization methods to decompose the rectangular-coordinate AC microgrid SLP formulation across weakly coupled subproblems like DERs or multi-microgrids. Blocker: None
MICROGRID ENERGY MANAGEMENT SYSTEM WITH ANCILLARY SERVICES TO THE GRID · Georgia Tech
Left open
Develop a Koopman operator-based distributed energy resource (DER) control strategy for multi-generator power systems beyond a single wind farm. Blocker: None
Koopman operator based data-driven characterization and control of power system dynamics and stability · Iowa State
Left open
Design and validate a decentralized power supply architecture to eliminate single-point-of-failure power distribution in multicore embedded avionics clusters. Blocker: Requires physical hardware prototyping, power electronics design, and embedded testbench apparatus.
Optimization-based design of fault-tolerant avionics · Georgia Tech
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.