Local Model Information Research Articles

FedAFR: Enhancing Federated Learning with adaptive feature reconstruction

Federated learning is a distributed machine learning method where clients train models on local data to ensure that data will not be transmitted to a central server, providing unique advantages in privacy protection. However, in real-world scenarios, data between different clients may be non-Independently and Identically Distributed (non-IID) and imbalanced, leading to discrepancies among local models and impacting the efficacy of global model aggregation. To tackle this issue, this paper proposes a novel framework, FedARF, designed to improve Federated Learning performance by adaptively reconstructing local features during training. FedARF offers a simple reconstruction module for aligning feature representations from various clients, thereby enhancing the generalization capability of cross-client aggregated models. Additionally, to better adapt the model to each client’s data distribution, FedARF employs an adaptive feature fusion strategy for a more effective blending of global and local model information, augmenting the model’s accuracy and generalization performance. Experimental results demonstrate that our proposed Federated Learning method significantly outperforms existing methods in variety image classification tasks, achieving faster model convergence and superior performance when dealing with non-IID data distributions.

Online model composition for networked discrete-event systems

Networked discrete-event systems consist of physically coupled subsystems that utilise a digital communication network for solving their current tasks. Each subsystem has to be steered from its current state into a situation-dependent target state. Due to the physical couplings among the subsystems, certain state transitions have to be executed synchronously with other subsystems to reach the local target states. To satisfy all local tasks, the subsystems have to find a deadlock-free order of the synchronous state transitions by applying global model information. However, in the networked discrete-event system there does not exist any coordinator and the global model information is distributed among the subsystems. Hence, the subsystems have to exchange model information and to compose it online to find the deadlock-free order of synchronous state transitions. This paper proposes an online model composition algorithm based on the A⁎-Algorithm that determines the necessary local model information to be communicated and composed by the subsystems. The method is demonstrated by an example.

Incremental Approximate Dynamic Programming for Nonlinear Adaptive Tracking Control with Partial Observability

Approximate dynamic programming is a class of reinforcement learning, which solves adaptive, optimal control problems and tackles the curse of dimensionality with function approximators. Within this category, linear approximate dynamic programming provides a model-free control method by systematically using a quadratic cost-to-go function. Although efficient, linear approximate dynamic programming methods are difficult to apply to nonlinear systems or time-varying systems. To overcome the above limitations, this paper proposes an adaptive nonlinear tracking control method based on incremental approximate dynamic programming, which combines the advantages of linear approximate dynamic programming and incremental nonlinear control techniques. This is a model-free method for unknown, nonlinear systems and time-varying references. The trait of separating the local model information from the cost function approximation makes this method an option for partially observable control problems. This paper, therefore, proposes two reference tracking controllers for different observability conditions: the direct measurement of the full state, and the partially observable tracking error. In each condition, two algorithms are developed for off-line learning and online learning, respectively. These algorithms are applied to attitude control of a spacecraft disturbed by internal liquid sloshing. The results demonstrate that the proposed algorithms accurately deal with the unknown, time-varying internal dynamics while retaining efficient control, even with only partial observability.

Model based urban traffic control, part I: Local model and local model predictive controllers

This paper is the first in a series of reports presenting a framework for the hierarchical design of feedback controllers for traffic lights in urban networks. The goal of the research is to develop an easy to understand methodology for designing model based feedback controllers that use the current state estimate in order to select the next switching times of traffic lights. In this paper we introduce an extension of the cell transmission model that describes with sufficient accuracy the major causes of delay for urban traffic. We show that this model is computationally fast enough such that it can be used in a model predictive controller that decides for each intersection, taking into account the vehicle density as estimated along all links connected to the intersection, what switching time minimizes the local delay for all vehicles over a prediction horizon of a few minutes. The implementation of this local MPC only requires local online measurements and local model information (unlike the coordinated MPC, to be introduced in the next paper in this series, that takes into account interactions between neighbouring intersections). We study the performance of the proposed local MPC via simulation on a simple 4 by 4 Manhattan grid, comparing its delay with an efficiently tuned pretimed control for the traffic lights, and with traffic lights controlled according to the max pressure rule. These simulations show that the proposed local MPC controller achieves a significant reduction in delay for various traffic conditions.

A Stealth Integrity Targeted Cyber‐Attack in Distributed Electric Power Networks with Local Model Information

AbstractElectric power networks are critical infrastructures, and their correct operation is of vital importance. Nowadays, these systems are prone to cyber‐attacks because of new vulnerabilities in the system and access to shared networks. In this paper, a novel Stealth Integrity Targeted Attack (SITA) is proposed in the context of distributed power systems. A distributed power system comprises several sub‐networks, or zones with dedicated control and monitoring centers. The overall system is represented by linear time invariant state space models with coupled dynamical and algebraic equations. In the proposed strategy, the attacker has access to only one of the sub networks; therefore, the attacker only requires local information about one of the power system zones. Primarily, the proposed attack policy is defined based on zero‐dynamics of the sub network. The intruder injects predesigned signals to both the local generation unit controller as well as local unsecured and controllable loads in the attacked zone. Moreover, the local measurement system, or the sensors of the targeted zone are tampered. Furthermore, it will be proved that although the neighbor zones have physical connections with the attacked zone, the injected adversary signals are designed as they do not impact other zones directly in order to conceal the local attack from neighbor control centers as much as possible. We provide some advice to system administrators to make the intrusion unfeasible or to reveal the attack. The simulations on IEEE‐118 bus test system illustrate the validity of the assertions.

Distributed control design with local model information and guaranteed stability

Most results on distributed control design of large-scale interconnected systems assume a central designer with global model knowledge. The wish for privacy of subsystem model data raises the desire to find control design methods to determine an optimal control law without centralized model knowledge, i.e. in a distributed fashion. In this paper we present a distributed control design method with guaranteed stability to minimize an infinite horizon LQ cost functional. The introduction of adjoint states allows to iteratively optimize the feedback matrix using a gradient descent method in a distributed way, based on a finite horizon formulation. Inspired by ideas on stabilizing model predictive control, a terminal cost term is used, which gives a bound on the infinite horizon cost functional and ensures stability. A method is presented to determine that term in a distributed fashion. The results are validated using numerical experiments.

Optimal structured static state-feedback control design with limited model information for fully-actuated systems

We introduce the family of limited model information control design methods, which construct controllers by accessing the plant’s model in a constrained way, according to a given design graph. We investigate the closed-loop performance achievable by such control design methods for fully-actuated discrete-time linear time-invariant systems, under a separable quadratic cost. We restrict our study to control design methods which produce structured static state feedback controllers, where each subcontroller can at least access the state measurements of those subsystems that affect its corresponding subsystem. We compute the optimal control design strategy (in terms of the competitive ratio and domination metrics) when the control designer has access to the local model information and the global interconnection structure of the plant-to-be-controlled. Finally, we study the trade-off between the amount of model information exploited by a control design method and the best closed-loop performance (in terms of the competitive ratio) of controllers it can produce.

A Scalable Modularized Synthesis Method for Distributed Kalman Filters

AbstractThis paper presents a scheme to construct distributed observers for a system consisting of agents interconnected in a graph structure. The scheme is an iterative procedure to improve the observers with respect to global performance. It is modular in the sense that each agent iterates using only local model information. As a consequence, the complexity of the scheme scales linearly with the size of the system. The resulting observers estimate states for each agent using only local measurements and model knowledge of its neighbors.Distributed observers are suboptimal to centralized ones and it is desirable to measure the amount of performance degradation. We show how to use the variables of the synthesis scheme to also determine such a measure of the suboptimality.

Distributed Estimation and Detection under Local Information

This work considers the problem of obtaining optimal estimates via distributed computation in a large scale system. The electric power system, the transportation system, and generally any computer or network system, are examples of large scale systems: a decentralized estimation of signals based on observations acquired by spatially distributed sensors is the basis for a wide range of important applications. In this work, we focus on the problem of reconstructing the initial state of a linear network in the presence of process and measurement noise. We consider a local model information setup, in which the entire dynamical and measurement model is nowhere available and cannot be reconstructed for the computation. Our estimation procedure relies upon a novel technique to solve a consistent system of linear equations, for which we prove correctness and convergence. In the second part of the paper we consider the problem of detecting anomalies in a large scale network driven by noise. Despite the theoretical advances in this field of research, the currently available procedures to enforce security in large scale systems are computationally inefficient and numerically unreliable. Using our optimal estimation scheme, we describe a distributed procedure with performance guarantees that only requires local knowledge of the system model.