Research On Key Points Of Multimedia Content Delivery For Machine Decision-Making

The rapid development of multi-modal artificial intelligence and the ubiquitous trend of computing resources indicate that,Internet of Things(Io T)will evolve into an intelligent system which is entirely based on machine decision-making.That is,the distributed multimedia sensors continuously acquire information of physical environments,and deliver the multimedia observations to resource-rich servers.Through processing these sensed content based on multi-modal artificial intelligence,the machine will independently make decisions without human intervention.Artificial intelligence of things,which combines multi-modal artificial intelligence and wireless content delivery,has become a key enabler for intelligent upgrading in traditional industries,and has given birth to a series of emerging applications such as smart home,intelligent transportation,and intelligent manufacturing.Nevertheless,building the intelligent Io T highly depends on the paradigm shift of multimedia content delivery from serving user experience to serving the training and inference of artificial intelligence,which incurs novel technical challenges.In terms of communication quality evaluation,machine intelligence has replaced human beings as the main consumer of multimedia content delivery,and significantly distinguishes in perceiving the distribution of key information than that of carbonbased brain.Moreover,multimedia content is no longer pushed from the centralized content server to the distributed clients for human viewing,but converges from heterogeneous sensing devices to the computing server for model training.As for network resource allocation,machine decision-making oriented content delivery systems rely on the collaborative design of sensing,communication,computation functionalities as well as training strategies of artificial intelligence,rather than optimizing the communication framework alone.Given the above paradigm shifts faced by multimedia content delivery for enabling machine decision-making,the main contributions and innovations of this thesis are summarized as follows:1)To save transmission costs for computation offloading,a machine decision-making oriented communication quality evaluation index is proposed.This index aims to avoid over provisioning of communication quality for serving intelligent applications.First,the functionality of physical components and logical processes affecting machine decision-making are carefully studied.Second,through jointly modeling the factors of commonness and individuality,a layered evaluation framework is designed,which is capable of monitoring the quality of data to acquired,delivered,and processed for machine decision-making.Besides,the thesis puts forward the design rationales and use cases of video transmission schemes when serving machine decision-making,i.e.,how to reduce communication quality provisioning without losing inference accuracy.Extensive experiments for intelligent road traffic monitoring demonstrate that,exploiting the differences between human-centric and machine-centric attentions of data distortion,has a unique advantage in saving transmission costs.2)To enhance the diversity and converage of multimedia content for the model training of artificial intelligence,a crowdsensing-based content sharing mechanism is proposed.The proposed mechanism allows participants to customize their own demands for crowdsensing,and introduces social network effects to make up for the enthusiasm loss derived from personalized demands.The strategic interactions between crowdsensing organizers and potential content providers,are characterized by a two-stage supermodular game.A fair participation mechanism and an optimal pricing strategy are developed,by solving the Pareto-optimal subgame-perfect Nash equilibrium of the proposed game with and without social attributes.The effectiveness of the proposed mechanism in stimulating content sharing,is verified by extersive simulations based on the Erdos-Renyi graph model and real-world dataset.3)To optimize the performance trade-offs between data acquisition,delivery,and processing,a novel radio resource allocation scheme is proposed for integrated sensing,communication,and computation networks.It focuses on avoid inefficient competition among different tasks for multidimensional resources from a perspective of joint optimization.First,a system model is established,where the integrated sensing and communication technology and edge computing paradigm is combined for enabling intelligent Io T systems.Then,a joint user association and channel assignment problem is formulated to capture the intrinsic conflicts among heterogensous requests.Due to its intractability,this resource allocation problem is reformulated as a two-sided matching problem with externality.An iterative matching algorithm is developed by introducing pairwise stability and proved to be convergent and stable.The simulation results elucidate the significant superiority of the proposed scheme over those externality-unaware wireless schedulers in load balancing.4)To utilize the idle communication and computation resources derived from the asynchronous task arrivals,this thesis proposes a hierarchical task offloading framework consisting of Device-toDevice computation offloading,edge computing,and cloud computing,in which cooperative relay technology is introduced to assist devices to access remote computing resources.Based on the analysis toward latency and energy consumption induced by limited edge computing capability,radio reuse interference,and backhaul network congestion,a sequential game based mode selection strategy is developed for the multi-user scenarios of the proposed framework.The simulation results demonstrate that the proposed strategy can effectively reduce end-to-end offloading latency and energy consumption,and also has good stability even when the network resources are limited.