Goal-Oriented Orchestration of Communication and Computing in Collaborative Intelligence Systems
Abstract
Collaborative Intelligence (CI) leverages distributed resources across interconnected devices to enhance decision-making capabilities. Propelled by rapid advancements in Artificial Intelligence and communication technologies, CI holds great promises for enabling diverse vertical applications, such as Distributed Machine Learning (DML) and intelligent transportation systems. While CI reduces reliance on centralized servers, it necessitates a paradigm shift toward integrated communication and computing, which raises a critical question: How do we orchestrate communication and computing to achieve seamless synergy, rather than treating them as loosely connected components? This challenge stems from the mismatch between application-layer demands, such as real-time processing, and physical-layer constraints, including resource availability, and leads to several key bottlenecks: resource contention from multidimensional demands, inefficiencies due to overlooked dependencies in integrated tasks, conflicts between global and local objectives driven by heterogeneity, and inadaptivity to dynamic conditions. To overcome these, this thesis proposes goal-oriented orchestration strategies of communication and computing for effective, efficient and adaptive CI systems.
Initially, to address overlooked dependencies, a concurrent communication-dependent computing task mechanism is introduced for DML applications, which models multidimensional requirements and defines a utility function to quantify the dependencies between communication and computing. By optimizing the utility, we formulate a task orchestration and resource management problem and solve it using auxiliary graphs and multi-agent reinforcement learning to deal with incomplete system information concerns in distributed decision-making.
Moreover, to accommodate evolving objectives and conditions in split learning, a hypergraph-aided model splitting mechanism is proposed to coordinate coupled tasks, including data preprocessing, model partition and resource management. Hypergraphs model the higher-order dependencies between performance objectives and coordination schemes, reducing computational complexity. A meta reinforcement learning algorithm further enables rapid adaptation to varying resource conditions and objectives, reducing frequent retraining.
Furthermore, to balance global-local objectives in federated learning, we design an adaptive federated meta learning framework that dynamically integrates heterogeneous device-specific conditions (computation/data/communication) into multimodal learning, to ensure global model accuracy and time cost, while enabling local adaptation via meta parameters. Similarly, for heterogeneous vehicles in collaborative perception, their locally extracted features can be spatiotemporally misaligned due to unsynchronized clocks, mobility, etc., degrading fusion accuracy and resource efficiency. We overcome these by proposing a feature-oriented resource management mechanism based on spatio-temporal alignment for collaborative perception. Simulations show significant improvements in the accuracy and efficiency of collaborative perception.