1,721,060 research outputs found
A Distributed Framework for k-hop Control Strategies in Large-Scale Networks Based on Local Interactions
No full textIn this paper, we propose a distributed framework for large-scale networks to attain control strategies requiring k-hop interactions. This research is motivated by the observation that in many practical applications and operational domains involving large-scale networks, such as environmental monitoring or traffic load balancing, agents may be required to collect only information concerning other agents located sufficiently close to them, that is agents topologically at most k-hop away. In this setting, distributed observers available at the state of art, which typically estimates the full network state, may be inadequate due to scalability issues. Differently, we propose a distributed finite-time observer which allows each agent to estimate the state of its k-hop neighbors by interacting only with the agents belonging to its 1-hop neighborhood. Furthermore, we demonstrate that for feedback control strategies based on k-hop neighborhood information, which are input-to-state stable, the proposed distributed finite-time observer can be effectively used to design stable large-scale networked control strategies. Numerical results are provided to corroborate the theoretical findings
A k-hop graph-based observer for large-scale networked systems
No full textIn this paper, we address the state estimation problem for multi-agent systems interacting in large scale networks. This research is motivated by the observation that in large-scale networks for many practical applications and domains, each agent only requires information concerning agents spatially close to its location, let's say topologically k-hop away. We propose a scalable framework where each agent is able to estimate in finite-time the state of its k-hop neighborhood by interacting only with the agents belonging to its 1-hop neighborhood
Evolutionary and Iterative Training of Recurrent Neural Networks via the Singular Value Decomposition
Full text linkLa tesi esamina l'uso della decomposizione ai valori singolari (SVD) dall'algebra lineare come strumento per l'analisi delle reti neurali, nonché il suo utilizzo per accelerare o addirittura limitare l'apprendimento (ad esempio, per prevenire l'over-fitting o mantenere la stabilità) e come base per gli algoritmi di apprendimento iterativo ed evolutivo.
Ciò che si presenta sono metodi per tenere conto della struttura intrinseca della trasformazione, anche durante l'utilizzo di metodi evolutivi, impiegando la decomposizione ai valori singolari. Naturalmente, il tentativo di preservare una certa struttura delle trasformazioni non è inedito, sia che questo significhi preservare la scarsità sia che si riferisca a qualche tipo di invarianza, come nell'invarianza di spostamento di uno strato convoluzionale.
I metodi presentati nel lavoro consentono di addestrare reti neurali ricorrenti per una varietà di problemi con cambiamenti nel tempo, tra cui la previsione dei prezzi, la manutenzione predittiva, l'identificazione del modello e il controllo automatico. Il nostro metodo non si basa sulla propagazione all'indietro e può essere utilizzato in ambienti supervisionati o non supervisionati. Inoltre, i nostri modelli possono essere facilmente inizializzati utilizzando la conoscenza del dominio o il metodo dei minimi quadrati (lineari) per "pre-programmare" il modello e iniziare l'ottimizzazione in un'area dello spazio della soluzione suscettibile di produrre risultati. Infine, data una rete neurale precedentemente addestrata in un dominio, i nostri modelli e metodi consentono il riutilizzo e la rapida riqualificazione per un dominio simile, preservando la struttura intrinseca della trasformazione nel cuore della rete neurale.This work examines the use of the singular value decomposition (SVD) from linear algebra as a tool for the analysis of neural networks, as well as its use to speed up or even limit learning (to prevent over-fitting or maintain stability, for example) and as the basis for iterative and evolutionary learning algorithms.
What we present here are methods of taking the inherent structure of the transformation into account — even while using evolutionary methods — using the singular value decomposition. Of course, preserving some structure of the transformations is not completely new — whether this means preserving sparseness or some type of invariance, as in the shift invariance of a convolutional layer.
The methods we present allow us to train recurrent neural networks for a variety of problems with changes through time, including price prediction, predictive maintenance and model identification, and automatic control. Our method does not rely on back propagation and can be used in either supervised or unsupervised settings. Further, our models can be easily initialized by using either domain knowledge or (linear) least squares to “pre-program” the model and begin optimization in an area of the solution space likely to yield results. Finally, given a neural network previously trained in one domain, our models and methods allow the reuse and quick retraining for a similar domain, by preserving the inherent structure of the transformation at the heart of the neural network
Dynamic Resilient Containment Control in Multirobot Systems
Full text linkIn this article, we study the dynamic resilient containment control problem for continuous-time multirobot systems (MRSs), i.e., the problem of designing a local interaction protocol that drives a set of robots, namely the followers, toward a region delimited by the positions of another set of robots, namely the leaders, under the presence of adversarial robots in the network. In our setting, all robots are anonymous, i.e., they do not recognize the identity or class of other robots. We consider as adversarial all those robots that intentionally or accidentally try to disrupt the objective of the MRS, e.g., robots that are being hijacked by a cyber–physical attack or have experienced a fault. Under specific topological conditions defined by the notion of (r,s)-robustness, our control strategy is proven to be successful in driving the followers toward the target region, namely a hypercube, in finite time. It is also proven that the followers cannot escape the moving containment area despite the persistent influence of anonymous adversarial robots. Numerical results with a team of 44 robots are provided to corroborate the theoretical findings
Secure rendezvous and static containment in multi-agent systems with adversarial intruders
No full textMulti-Agent Coordination of Thermostatically Controlled Loads by Smart Power Sockets for Electric Demand Side Management
No full textThis article presents a multi-agent control architecture and an online optimization method based on a dynamic average consensus to coordinate the power consumption of a large population of thermostatically controlled loads (TCLs). Our objective is to penalize peaks of power demand, smooth the load profile, and enable demand-side management. The proposed architecture and methods exploit only local measurements of power consumption via smart power sockets (SPSs) with no access to their internal temperature. No centralized aggregator of information is exploited, and agents preserve their privacy by cooperating anonymously only through consensus-based distributed estimation. The interactions among devices occur through an unstructured peer-to-peer (P2P) network over the Internet. Methods for parameter identification, state estimation, and mixed logical modeling of TCLs and SPSs are included. The architecture is designed from a multi-agent and plug-and-play perspective, in which existing household appliances can interact with each other in an urban environment. Finally, a novel low-cost testbed is proposed along with numerical tests and experimental validation
A Heuristic approach for Online Distributed Optimization of Multi-Agent Networks of Smart Sockets and Thermostatically Controlled Loads based on Dynamic Average Consensus
No full textThis paper presents a novel heuristic online optimization method and multi-agent control architecture to optimize the Peak-to-Average power Ratio (PAR) of a large population of Thermostatically Controlled Loads (TCLs) over a sliding receding horizon time window. The proposed architecture exploits only local measurements of the TCL power consumption with no knowledge of their internal temperature. No centralized aggregator of information is used and agents preserve their privacy by cooperating only through consensus-based distributed estimation. TCLs interactions occur via Smart Power Sockets (SPSs) which are interconnected through a peer-to-peer (P2P) network over the internet. The control architecture is designed from a multi-agent perspective in which real household appliances can interact with each other via SPSs.Our contribution is twofold: first we introduce a novel hybrid modelling of the TCL-plus-SPS system along with a method for parameter identification and a method estimate the internal state of the TLC through SPS performed power measurements; then we provide a heuristic algorithm for online distributed optimization of the on/off states of the SPSs which exploits a dynamic average consensus algorithm to estimate the planned future average power consumption of the network while preserving the agents' privacy. Numerical simulations and preliminary experimental results performed in a novel low cost testbed are provided
Robust Containment Control in Multi-Agent Systems with Common Coordinate Frames and Bearing Angle Measurements
No full textIn this paper, the robust containment control problem in multi-agent systems (MASs) with multiple static leaders and with malicious agents is addressed. In our setting, we define as malicious all those agents which do not implement the local control protocol executed by the followers in the MAS. On the contrary, we assume malicious agents apply a control input of their own choice with the intent of jeopardizing the cooperation in order to bring the followers arbitrarily away from a containment area, i.e., an hypercube defined by the location of the leaders. For this setting, a distributed protocol, which is proven to be robust against malicious agents under certain topological conditions, is considered. It is assumed that the agents move in a d-dimensional hyperplane, share a common coordinate system, do not require access to absolute positions (GPS) and are able to measure bearing angles of their neighbors. A theoretical characterization of the proposed algorithm is provided together with numerical results
- …
