Emmanouel Varvarigos

Application placement across the multi-tenant edge–cloud continuum remains suboptimal, as various providers struggle to coordinate provisioning for efficient resource access across administrative domains. Resource federations can mitigate this fragmentation by enabling providers to pool heterogeneous resources and jointly orchestrate their allocation. However, modern applications' requirements cannot be addressed through small-scale unification alone and require provisioning across multiple federations. In this work, we consider a multi-federation computing substrate and introduce a coordination framework for cross-federation inter-operation. We formulate the problem as a prosumer market where federations are incentivized to dynamically procure or lease out resources depending on demand and scarcity conditions. Coordination is realized through a dynamic multi-dimensional reverse auction that jointly determines the offered resource configuration and its lease price. The framework accounts for competition at both the federation and constituent provider levels, jointly optimizing latency, federation profitability and equitable profit distribution. Market interactions are formulated as a Partially Observable Stochastic Game with federations as self-interested agents. To handle the high-dimensional action space, we introduce MA3PO-SAC, a Multi-Agent Hierarchical Reinforcement Learning architecture that decomposes bidding into specialized networks for discrete resource selection and continuous pricing. Trace-driven experiments using the Google Borg trace and cloud-calibrated infrastructure configurations demonstrate that MA3PO-SAC learns adaptive bidding attuned to market conditions and to both individual and peer objectives. The method outperforms alternative MARL baselines and achieves system welfare within 3\% of a centralized full-visibility orchestrator, despite partial observability, inter-federation competition, and decentralized control.

Cloud-native applications, increasingly structured as chains of interconnected microservices, improve performance by distributing components across the edge–cloud continuum. However, emerging edge–cloud resource Associations comprise heterogeneous administrative domains operated by distinct tenants. Deploying services across such multi-tenant environments with varying levels of inter-domain trust introduces significant orchestration complexity and raises critical security challenges. This work proposes a holistic resource orchestration framework that jointly addresses execution-plane and communication-plane security. We investigate graded service–host isolation using state-of-the-art lightweight virtualization technologies (including generic containers, software and hardware sandboxes, and unikernels) while explicitly accounting for their distinct CPU footprints and operational costs. In parallel, we incorporate mutual Transport Layer Security (mTLS) for inter-service communication and mathematically model its associated computational and latency overheads. We formulate the resulting orchestration problem as a multi-objective Integer Linear Program (ILP) that optimizes a weighted combination of monetary execution cost, end-to-end latency, and security. To tackle its complexity, we introduce a multi-agent Rollout algorithm based on approximate dynamic programming, which systematically improves a myopic best-fit placement policy. Extensive simulations demonstrate that: (i) the proposed Rollout mechanism significantly improves placement quality, reducing the optimality gap to below 7.5% across all optimization scenarios while maintaining high cosine similarity with optimal solver’s decisions; (ii) explicitly modeling security decisions reveals important operational trade-offs; and (iii) evaluating orchestration under isolated, zero-trust, federated, and full-trust environments shows that even partial inter-tenant trust substantially improves performance compared to zero-trust deployments, while zero-trust orchestration still provides notable benefits over completely isolated infrastructures.

User Equipment (UE)-centric Cell-Free massive Multiple-Input Multiple-Output (CF mMIMO) has emerged as a key enabling architecture for sixth-generation (6G) networks, offering a favorable trade-off between performance gains, cost and energy-efficiency. However, optimizing Access Point (AP) deployment and hardware provisioning in UE-centric CF mMIMO remains a challenging task, because planning decisions are tightly coupled with complex operational decisions such as the dynamic AP-UE clustering. In this paper, we address the joint planning-and-operation problem for UE-centric CF mMIMO networks under realistic cost, energy and performance-related models and constraints. We propose a hierarchical optimization framework in which a novel Deployment Evolutionary Algorithm (DEA) iteratively refines AP locations and heterogeneous hardware configurations, while a Weight-driven Greedy Round-Robin (WG-RR) algorithm acts as an operational oracle, performing UE-centric AP–UE clustering during the operational phase in a manner consistent with operator-defined business intents (i.e. objective weights). To ensure robust and realistic network dimensioning, candidate deployments are evaluated over multiple traffic demand-oriented snapshots that capture UE mobility and spectral-efficiency demand fluctuations. Extensive multi-objective and multi-snapshot evaluations reveal significant spectral-efficiency diminishing returns with increasing deployment and cooperation density, exposing practical operating points that substantially reduce CAPEX and operational power with negligible performance loss. The results further demonstrate that heterogeneous AP deployments enable superior planning decisions, thereby supporting the gradual and sustainable integration of UE-centric CF mMIMO into existing multi-cell infrastructures.

Demand Aggregators (DAs) are expected to have a continuously increasing impact in the operation of modern electricity markets. Such players tend to concentrate large portfolios of demand-side resources, leading to market power concentration and imperfect competition. Such imperfectly competitive markets are analyzed in the literature by studying their equilibria. However, little attention has been given to the learning dynamics through which players adapt their bidding strategies over time and, crucially, to which equilibria are actually learnable. In this paper, we propose the coarse-correlated equilibrium as a solution concept that is more relevant than the Nash equilibrium for such markets. In contrast to the typical approach of the literature on Equilibrium Problems with Equilibrium Constraints, we model the learning dynamics that strategic DAs use to gradually optimize their bidding strategies. At the same time, in contrast to generic model-free approaches, we account for the fact that market players would make better use of the markets’ available information by explicitly modeling their competitors. To that end, we propose a game-theoretic learning algorithm, in which each strategic DA observes the bids of its competitors and gradually builds beliefs about their bidding strategies. The resulting equilibria are compared to the pure Nash equilibria of the related literature. Our numerical results demonstrate that, in practice, strategic DAs can manipulate prices and market conditions to a greater extent than previously anticipated for the benefit of the energy system as a whole.

Managing resource use in cloud and edge environments is crucial for optimizing performance and efficiency. Traditionally, this process is performed with detailed knowledge of the available infrastructure while being application-specific. However, it is common that users cannot accurately specify their applications’ low-level requirements, and they tend to overestimate them—a problem further intensified by their lack of detailed knowledge on the infrastructure’s characteristics. In this context, resource orchestration mechanisms perform allocations based on the provided worst-case assumptions, with a direct impact on the performance of the whole infrastructure. In this work, we propose a resource orchestration mechanism based on intents, in which users provide their high-level workload requirements by specifying their intended preferences for how the workload should be managed, such as prioritizing high capacity, low cost, or other criteria. Building on this, the proposed mechanism dynamically assigns resources to applications through a Reinforcement Learning method leveraging the feedback from the users and infrastructure providers’ monitoring system. We formulate the respective problem as a discrete-time, finite horizon Markov decision process. Initially, we solve the problem using a tabular Q-learning method. However, due to the large state space inherent in real-world scenarios, we also employ Deep Reinforcement Learning, utilizing a neural network for the Q-value approximation. The presented mechanism is capable of continuously adapting the manner in which resources are allocated based on feedback from users and infrastructure providers. A series of simulation experiments were conducted to demonstrate the applicability of the proposed methodologies in intent-based resource allocation, examining various aspects and characteristics and performing comparative analysis.

This paper addresses the problem of efficiently scheduling of EV charging requests in a single Charging Station (CS), also taking into consideration the inability of EV users to express their preferences in closed-form utility functions. We consider a paradigm where a menu of possible charging intervals, each associated with an urgency price, is generated online. By letting the EV users pick their most preferable menu option, the proposed algorithm commits on each EV’s charging completion time upon its arrival, achieves a near optimal total weighted charging completion time, and prevents the users from strategically misreporting their preferences, while offering a practical and implementable solution to the problem of EVs -charging station interaction. Our experimental evaluation demonstrates that our algorithmic approach performs nearly optimal in a variety of practically relevant scenarios.