Research On Consensus And Differential Privacy Based On Event-triggered Control And Quantized Communication In Multi-agent Networks

Recent years,consensus and cooperation problems of multi-agent networks have attracted increasing attention from various fields such as multi-robot coordination,sensor fusion,objective tracking,filtering and estimation,large-scale machine learning,distributed computation and optimization.The consensus problems mainly focus on studying how a global network can achieve an agreement by using local available information based on local communication.That is,each agent in a network does not need to communicate with all other agents and only exchange state with its neighbors to update its own state such that the global consensus of a network can be achieved.Though there are a number of works on consensus study,the existing works disregard many important factors in practical implementation.For example,most existing works all assume that each agent can send and receive exact real-value state with no errors,which is quite unrealistic because some digital devices such as analog-to-digital and digital-to-analog convertors,discrete-level actuators and digital communication channels are often embedded in a real digital network.Also,there may exist frequent real-time communication and controller updates in the existing works.In addition,from the perspective of privacy,the participating agents may not want to disclose their own states while communicating with each other to reach an agreement.Thus,the privacy-preserving problem in the consensus study of multi-agent networks becomes another challenge for researchers.Recent years,the quantized communication technology converting analog signals to digital signals is an important topic in the study of consensus due to the significant advantages of digital signals on robustness and security compared with analog signals.To avoid frequent real-time communication and controller updates,the event-triggered control mechanism offers a new point of view on how information should be collected and transmitted.Based on the event-triggered control strategy,an agent does not need to transmit its local state at every time instant,instead,only transmits its local state to its neighbors only when it is necessary,that is,only when a measurement of the local state error reaches a specified threshold.In the context of privacy preservation,the differential privacy has gained significant popularity due to its rigorous formulation and proven security properties,including the resilience to post processing and side information,and the independence from the models of adversaries.Thus,it is interesting to introduce the notion of the differential privacy to the consensus study.The main contribution of this dissertation is that it focuses on the consensus study by introducing the quantized communication technique,the event-triggered control mechanism and the differential privacy mechanism.The main points of this dissertation are summarized as follows:First,we briefly introduce basic average consensus models with the distributed event-triggered mechanism and the differential privacy mechanism.Under the event-triggered mechanism,all agents in a network can asymptotically achieve an average consensus with sporadic communication and less controller updates.Under the differential privacy mechanism,all agents in a network not only can select their own privacy level to prevent information disclosure independently but achieve a mean square average consensus with a bounded disturbance variance asymptotically.Second,we study the consensus problem of digital multi-agent networks by employing a dynamic encode/decode technology and a distributed event-triggered control strategy.We first design a novel integrated communication framework to describe the communication process in a digital channel.This communication framework employs a dynamic encode/decode technology and an event-triggered control strategy,which are responsible for dealing with the quantized communication and avoid frequent real-time communication and controller updates.According to this framework,a formal and detailed communication algorithm is provided in the integrated communication environment.Then,the quantized consensus model with the event-triggered control strategy is proposed and a novel distributed event-triggering condition is developed.Note that this event-triggering condition only depends on the available local information,which can effectively avoid the collection of the global information.Furthermore,the consensus analysis is given to explain the sufficiency and reasonability of the proposed results.In addition,a self-adaptive avoidance algorithm for the dynamic quantization factor is proposed,which effectively avoid the saturation of the quantizers.Third,we study the differentially private consensus problem of multi-agent networks by employing a distributed event-triggered mechanism such that not only the agents can protect the privacy of their initial states from information disclosure but the execution efficiency of the whole network can be improved.We first propose a distributed event-triggered mechanism for a differentially private consensus algorithm such that frequent real-time communication and controller updates can be avoided.Then,we propose a distributed event-triggering condition that only depends on local information and local parameters,which can effectively avoid global information collection.Furthermore,the convergence analysis of mean square average consensus is given to explain the sufficiency of the proposed event-triggered mechanism and event-triggering condition.In addition,we establish the statistic properties of the convergent accuracy that the expectation of the convergence point converges to the average value of all agents' initial states exactly and the disturbance variance is bounded with an explicit expression.Finally,we further give the differential privacy analysis that each agent can flexibly select its own privacy level to prevent information disclosure.Finally,we study the differentially private consensus problem under a quantized communication environment where the exact real-value state is not available for transmission due to the range limitation of digital channels.We first extend the differentially private consensus model to the case of a quantized communication environment integrated with a dynamic encoding/decoding scheme and propose a differentially private communication algorithm utilizing the quantized state with a bounded quantizer instead of the exact real-value state to reach an agreement while protecting the initial or current states of the participants from information disclosure.Then,the convergence analysis of mean square consensus in the case of an unbounded quantizer is given to explain the sufficiency of the extended model and convergence conditions.To overcome the uncertainty of saturation in the case of a bounded quantizer,we also give a statistical analysis on the boundedness of quantization that the bounded quantizer with a finite number of bits can remain unsaturated with desired high probability under certain conditions.Furthermore,we establish the statistical property of the convergent accuracy that the expectation of the convergence point converges to the average value of all agents' initial states exactly and that the disturbance variance is bounded with an explicit expression.In addition,we further discuss the differential privacy level from the view of privacy preservation and establish the relationship between the dynamic encoding/decoding scheme and the differential privacy mechanism.