A Multi-Agent Knowledge-Based System for Epistemic Uncertainty Management in Data Center Applications: Integrating Thermodynamic Optimization and Large Language Models for Reliable Operational Decision Support

Modern Data Center infrastructures operate at the intersection ofphysical uncertainty---thermal dynamics, hardware failure, powerfluctuations---and epistemic uncertainty: scarce incident history fornovel cooling technologies, ambiguous expert consensus duringhigh-pressure operational events, and semantically framed NOCcommunications that can bias human decision-making in criticalmoments. Classical probabilistic frameworks yield fragile estimateswhen incident logs are limited, while LLMs applied directly tooperational Decision Support Systems exhibit well-documented failuremodes: hallucinations, overconfidence, and susceptibility to theframing bias inherent in incident narratives and runbooks.We propose MAS-DSS, a hybrid multi-agentknowledge-based framework operationalising four uncertaintydimensions tailored to Data Center management --- Probability (\(P_1\),historical incident data), Plausibility (\(P_2\), Dempster--Shaferexpert consensus), Credibility (\(P_3\), NLP/sentiment analysis ofNOC communications and ITSM tickets), and Possibility (\(P_4\), fuzzylogic for Black Swan scenarios such as cascading cooling failures)--- fused into the Happenability metric (\(A \in [0,1]\)) viaa Dynamic Fusion Engine, complemented by Reliability (\(R\)) andinter-agent standard deviation (\(\sigma\)) as an epistemic qualityvector. The seven Critical Success Factors (CSFs) operationalisedinclude Energy Efficiency (PUE/WUE), Thermal Stability, InfrastructureReliability, Network Availability, Cybersecurity Posture, RegulatoryCompliance, and Operational Continuity. A per-agent segregated RAGpipeline (BGE-M3/ChromaDB) grounds each qualitative reasoning stepin verifiable domain knowledge prior to deterministic computation viathe Model Context Protocol, with LLMs acting exclusively assemantic orchestrators to structurally prevent hallucinations andanchoring bias.A thermodynamic optimisation engine (MMS-DSS) projects the decisionspace onto an Energy--Entropy plane \((E, S)\), bypassing POMDPintractability without explicit reward engineering, and navigatingoperational states from degradation conditions towardsDynamic Order \((E_f=28, S_f=0)\). A Joint Analysis layerintegrates both subsystems by projecting \((A, \sigma, R)\)along the optimal thermodynamic trajectory, yielding a step-by-stepinterpretable cognitivo-thermodynamic recommendation suitable forSite Reliability Engineers operating under real-time constraints.Validation demonstrates: formal framing invariance for CSF-anchoredagents (\(P_1\), \(P_4\)); graceful epistemic degradation under\(P_1\) knockout --- simulating the Fog-of-War conditions typical ofthe acute phase of a DC incident --- with categorical stability andaccurate quality signalling via (\(R\), \(\sigma\)); and structuralresistance to adversarial semantic framing --- three autonomous LLMsproduced categorical flips (\(\Phi_{\mathrm{LLM}} \in [0.17, 0.23]\))under positive framing (analogous to minimising incident descriptionsin NOC communications), while MAS-DSS maintained stability(\(\Phi_{\mathrm{MAS}} = 0.008\)). The architecture satisfies byconstruction the transparency and human oversight requirements ofthe EU AI Act and the EU Critical Entities Resilience Directivefor operators of essential infrastructure services.

Authors

Publication Details

Journal
Preprints.org
Published
2026-11-18
DOI
https://doi.org/10.20944/preprints202611.1886.v1
Primary Topic
Reliability and Maintenance Optimization
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Multi-Agent Knowledge-Based System for Epistemic Uncertainty Management in Data Center Applications: Integrating Thermodynamic Optimization and Large Language Models for Reliable Operational Decision Support

Marta Chinnici, Vincenzo Oliva, Salvatore Villano, Gerardo Iovane et al.
Preprints.org
Reliability and Maintenance Optimization
preprint

A Multi-Agent Knowledge-Based System for Epistemic Uncertainty Management in Data Center Applications: Integrating Thermodynamic Optimization and Large Language Models for Reliable Operational Decision Support

Marta Chinnici, Vincenzo Oliva, Salvatore Villano, Gerardo Iovane, Germano Ingenito
preprint en

Abstract

Modern Data Center infrastructures operate at the intersection ofphysical uncertainty---thermal dynamics, hardware failure, powerfluctuations---and epistemic uncertainty: scarce incident history fornovel cooling technologies, ambiguous expert consensus duringhigh-pressure operational events, and semantically framed NOCcommunications that can bias human decision-making in criticalmoments. Classical probabilistic frameworks yield fragile estimateswhen incident logs are limited, while LLMs applied directly tooperational Decision Support Systems exhibit well-documented failuremodes: hallucinations, overconfidence, and susceptibility to theframing bias inherent in incident narratives and runbooks.We propose MAS-DSS, a hybrid multi-agentknowledge-based framework operationalising four uncertaintydimensions tailored to Data Center management --- Probability (\(P_1\),historical incident data), Plausibility (\(P_2\), Dempster--Shaferexpert consensus), Credibility (\(P_3\), NLP/sentiment analysis ofNOC communications and ITSM tickets), and Possibility (\(P_4\), fuzzylogic for Black Swan scenarios such as cascading cooling failures)--- fused into the Happenability metric (\(A \in [0,1]\)) viaa Dynamic Fusion Engine, complemented by Reliability (\(R\)) andinter-agent standard deviation (\(\sigma\)) as an epistemic qualityvector. The seven Critical Success Factors (CSFs) operationalisedinclude Energy Efficiency (PUE/WUE), Thermal Stability, InfrastructureReliability, Network Availability, Cybersecurity Posture, RegulatoryCompliance, and Operational Continuity. A per-agent segregated RAGpipeline (BGE-M3/ChromaDB) grounds each qualitative reasoning stepin verifiable domain knowledge prior to deterministic computation viathe Model Context Protocol, with LLMs acting exclusively assemantic orchestrators to structurally prevent hallucinations andanchoring bias.A thermodynamic optimisation engine (MMS-DSS) projects the decisionspace onto an Energy--Entropy plane \((E, S)\), bypassing POMDPintractability without explicit reward engineering, and navigatingoperational states from degradation conditions towardsDynamic Order \((E_f=28, S_f=0)\). A Joint Analysis layerintegrates both subsystems by projecting \((A, \sigma, R)\)along the optimal thermodynamic trajectory, yielding a step-by-stepinterpretable cognitivo-thermodynamic recommendation suitable forSite Reliability Engineers operating under real-time constraints.Validation demonstrates: formal framing invariance for CSF-anchoredagents (\(P_1\), \(P_4\)); graceful epistemic degradation under\(P_1\) knockout --- simulating the Fog-of-War conditions typical ofthe acute phase of a DC incident --- with categorical stability andaccurate quality signalling via (\(R\), \(\sigma\)); and structuralresistance to adversarial semantic framing --- three autonomous LLMsproduced categorical flips (\(\Phi_{\mathrm{LLM}} \in [0.17, 0.23]\))under positive framing (analogous to minimising incident descriptionsin NOC communications), while MAS-DSS maintained stability(\(\Phi_{\mathrm{MAS}} = 0.008\)). The architecture satisfies byconstruction the transparency and human oversight requirements ofthe EU AI Act and the EU Critical Entities Resilience Directivefor operators of essential infrastructure services.

Preprints.org
Reliability and Maintenance Optimization
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