Back to search

Hellenic Open University

Collective Artificial Intelligence and Machine Learning with resource constraints

Abstract

dc:description

Trust constitutes a fundamental mechanism for decision making in open and dynamic Multi-Agent Systems, where autonomous agents must identify reliable partners under conditions of uncertainty, incomplete information, and continuously changing environments. Although computational trust and reputation models have been extensively studied and successfully applied in numerous domains, important challenges remain unresolved, including the cold-start problem, agent mobility, behavioral volatility, communication overhead, vulnerability to disinformation attacks, and privacy concerns, and resource constraints in terms of computational, communication, and decision-making costs. This dissertation investigates alternative approaches to trust management in dynamic Multi-Agent Systems and introduces a trustee-centric perspective on computational trust. Inspired by the biological mechanisms of synaptic plasticity and neural assembly formation, a novel trust model, named Create Assemblies (CA), is proposed. Unlike conventional trust-management approaches, where trustors evaluate and select trustees, the proposed model enables trustees to actively participate in the trust-formation process by assessing their own suitability for performing requested tasks. The proposed approach is evaluated through a series of simulation experiments and compared against FIRE, a well-known trust and reputation model for open Multi-Agent Systems. The results demonstrate that the trustee-centric approach exhibits distinct advantages under specific environmental conditions, particularly in scenarios characterized by changes in trustor populations, while conventional trustor-centric approaches remain advantageous in other situations. These findings indicate that no single trust-management mechanism consistently outperforms all alternatives across all operating conditions. To address this limitation, the dissertation further investigates adaptive trust management. First, a self-assessment mechanism is incorporated into the CA model, enabling trustees to detect changes in their own performance and adjust their behavior accordingly. Second, the problem of trust model selection is formulated as a reinforcement learning task. A Deep Q-Learning-based adaptable trustor is developed, capable of dynamically selecting between alternative trust mechanisms according to observed environmental conditions. Experimental results demonstrate that adaptive trust model selection can achieve robust performance across a wide range of dynamic environments and often achieves or exceeds the performance of the best individual trust model. Beyond performance evaluation, the dissertation examines the broader properties of the proposed trustee-centric paradigm, including privacy preservation, resistance to reputation manipulation attacks, communication efficiency, explainability, and scalability. Furthermore, the applicability of the approach beyond traditional trust-management problems is explored through the development of a simulation framework for emergency department operations, where trust-informed self-assessment and task allocation mechanisms are employed as decision-support tools. The findings of this dissertation suggest that trust management should not be viewed as a competition between alternative trust models seeking a universally superior solution. Instead, different trust mechanisms exhibit complementary strengths under different environmental conditions. The proposed trustee-centric approach provides a complementary perspective to conventional trustor-centric models, while reinforcement learning offers a mechanism for dynamically selecting the most appropriate trust strategy according to prevailing circumstances. Together, these findings contribute towards the development of adaptive, robust, and intelligent trust-management frameworks capable of operating effectively in complex, open, and continuously evolving environments under resource-constrained conditions where decision quality must be balanced against computational and communication cost.

Degree

thesis:*
Grantor dc:publisher
Hellenic Open University
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Lygizou, Zoi
  • Λυγίζου, Ζωή

Subjects

dc:subject × 29

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/62996

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Lygizou, Zoi; Λυγίζου, Ζωή. Collective Artificial Intelligence and Machine Learning with resource constraints. Hellenic Open University, 2026. http://hdl.handle.net/10442/hedi/62996