Research On Collaborative Computing For Edge Intelligence

With the integration of Artificial Intelligence(AI),Internet of Things(Io T),and Edge Computing,Edge Intelligence(EI)has gathered extensive attention in recent years.EI aims to push model training,model inference and intelligent decision from the cloud to the Internet edge,ca-pable of exploiting the widely distributed data,storage and computing re-sources therein to provide safe,timely,economical and reliable intelligent services.Albeit with its broad applications,since a single edge device typ-ically suffers from limited resources,achieving EI must rely on effective collaborative computing among devices to share the cost in model train-ing and inference,and improve the magnitude,dimension of knowledge required for AI.A significant body of research has been placing interests in collaborative computing in EI.However,it is still highly non-trivial to de-velop practical,efficient and safe collaborative computing strategies,due to the diversity of issues in EI and the complexity of edge networks.Building on the current research progress,this thesis attempts to systematically an-alyze the challenges faced by collaborative computing in EI,and carry out research from three aspects,i.e.,edge collaborative training,edge collab-orative inference and edge collaborative decision.The main contributions are outlined as follows:(1)Aiming at improving the personalization ability and communica-tion efficiency in edge-edge collaborative learning while facilitating its ef-fective deployment in edge networks,this thesis proposes a fast-convergent federated meta-learning(FML)algorithm along with a corresponding re-source allocation strategy in multi-access wireless networks.Specifically,this thesis rigorously provides a tight lower bound of the global loss re-duction in each round and quantifies the contribution of each device to the convergence.Based on the theoretical results,a new federated meta-learning algorithm(NUFM)is introduced,which can speed up the training by actively selecting participating devices in each round to maximize the global loss reduction in each round.To achieve effective deployment in edge networks,this thesis proposes a resource allocation strategy(URAL)to jointly optimize the convergence speed,wall clock time and energy con-sumption of NUFM in a multi-access wireless network.In particular,this thesis shows that the computational complexity of NUFM can be further reduced from O(d~2)to O(d)(d is the model dimension)by using two first-order approximation techniques.Extensive simulation results demonstrate the effectiveness and superiority of the proposed methods in comparison with existing baselines.(2)Aiming at facilitating fast and continual edge-edge collaborative learning,this thesis proposes a new federated meta-learning paradigm,which enables the end devices to learn on current tasks while exploiting the knowl-edge learned from previous tasks.Specifically,this thesis first formulates the edge-edge collaborative learning problem as a regularized optimiza-tion problem,where the Bregman divergence is used as a regularizer to facilitate knowledge transfer.To solve this problem,this thesis designs an ADMM-based federated meta-learning algorithm(ADMM-Fed Meta),where ADMM can decouple the original problem into multiple sub-problems that can be solved in parallel between end devices and the edge server.To further reduce the high computational cost in solving these sub-problems,this thesis devises an inexact-ADMM algorithm using first-order approxi-mation and Hessian estimation,which reduces the computational complex-ity in each round to O(n).In theory,this paper carried out a comprehensive analysis of ADMM-Fed Meta,in terms of convergence,fast adaptation,and forgetting effect.Extensive experimental studies demonstrate the effective-ness and efficiency of ADMM-Fed Meta and showcase that it substantially outperforms the existing baselines.(3)Aiming at facilitating effective model collaborative inference in edge,this thesis proposes a distributed online task offloading strategy with delay guarantees.This thesis first formulates the computation offloading problem as a delay-constrained long-term stochastic optimization problem under unknown prior statistical knowledge.To solve this optimization prob-lem,this thesis provides a technical path to transform and decompose the original problem into three slot-level subproblems.Based on the decom-position,this thesis develops an online algorithm(TODG)to solve each subproblem in a distributed manner.In particular,using the“δ-period strat-egy”,TODG only needs to allocate channels everyδslots,capable of re-ducing the computational cost largely.Theoretically,this thesis provides a comprehensive performance analysis of TODG,in terms of the optimality gap,the worst-case delay,and the impact of system parameters.Extensive simulation results demonstrate the effectiveness and efficiency of TODG.(4)Aiming at dealing with the cost and risk in the data collection of edge collaborative decisions,this thesis proposes two communication-efficient federated offline reinforcement learning(FORL)algorithms in no need of interaction with the environment.Using two novel regularization terms,this paper first proposes MF-FORL to approach the instability and inefficiency in FORL caused by intrinsic conservatism in offline reinforce-ment learning.Specifically,this thesis adds a value regularizer into the up-dates of local value functions to increase the values of good actions in the global data manifold.In addition,a policy regularizer is added to confine the local policy to the aggregated one,enabling it to utilize global informa-tion.The two regularization terms alleviate the overfitting problem during local updates and facilitate better offline policy evaluation,thereby signifi-cantly improving learning stability and communication efficiency.Further,this thesis develops a model-based approach(MB-FORL)that leverages a dynamic model estimated from the offline dataset.It enables the algorithm to learn on additional model-generated synthetic data to improve its gen-eralization ability.More importantly,using the learned model,the policy updating can be transferred from local devices to the server.The devices only need to evaluate the policy locally,thus largely reducing the computa-tional cost thereof.This thesis theoretically establishes the safe policy im-provement guarantees for MF-FORL,and evaluates the performance of the proposed algorithms on the standard offline RL benchmark.The extensive experimental results showcase that the proposed algorithms significantly outperform the baselines and enjoy great communication efficiency.(5)Aiming at tackling the cost of environmental interaction and the difficulty in manually designing reward functions when deploying decision algorithms in edge,this paper proposes a model-based offline inverse rein-forcement learning algorithm(CLARE),which can automatically infer the reward function in no need of interaction with the environment.In order to deal with the tricky reward extrapolation error caused by the intrinsic co-variate shift,CLARE properly incorporates conservatism into the learning reward using both(limited)expert data and(potentially sufficient)diverse data.Specifically,CLARE iterates between“conservative reward updat-ing”and“safe policy improvement”,and the reward function is updated by improving its values on weighted expert and diverse state-actions while cautiously penalizing those generated from model rollouts.As a result,it can encapsulate the expert intention while conservatively evaluating out-of-distribution state-actions,which in turn encourages the policy to visit data-supported states and follow expert behaviors and hence achieves safe policy search.The theoretical analysis provides an upper bound on the return gap between the learned policy and the expert policy,based on which we char-acterize the impact of covariate shift by examining subtle two-tier tradeoffs between the exploitation(on both expert and diverse data)and exploration(on the estimated dynamics model).We show that CLARE can provably alleviate the reward extrapolation error by striking the right exploitation-exploration balance therein.Extensive experiments corroborate the signifi-cant performance gains of CLARE over existing state-of-the-art algorithms.