Goal-Oriented Orchestration of Communication and Computing in Collaborative Intelligence Systems

#Goal-Oriented#Communications#Computing#Collaborative-Intelligence

Collaborative Intelligence (CI) leverages distributed data and computational resources across devices and/or servers, enhancing decision-making capabilities beyond those of isolated servers. Driven by rapid advancements in Artificial Intelligence (AI) and communication technologies, CI reduces reliance on centralized servers and shifts to a paradigm that integrates communication and computing. However, a significant question arises: How can we orchestrate communication and computing beyond mere connectivity? The challenge lies in the mismatch between application-layer demands, such as real-time processing, and physical-layer capabilities, including resource availability, and leads to critical bottlenecks: resource strain from multidimensional demands, inefficiency from overlooked task dependencies, conflicts between global and local optimization goals, and limited adaptability to changing conditions.

This thesis aims to create goal-oriented orchestration strategies for communication and computing to facilitate efficient, adaptive CI systems. To account for the overlooked dependencies, we introduce a concurrent communication-dependent computing task framework that models multidimensional requirements and defines a utility function to quantify the dependency impacts. By optimizing overall utility, we tackle task orchestration and resource management using auxiliary graphs and multi-agent reinforcement learning, enabling distributed decision-making under partially observable conditions.

To adapt to evolving objectives and conditions in split learning, a hypergraph-based dynamic model splitting mechanism is introduced, where the coupled task coordination and their goal achievement are modeled. It then enables rapid adaptation using a meta reinforcement learning algorithm that minimizes retraining overhead. Simulations under varying objectives and resource conditions demonstrate the effectiveness of our method.

For balancing global and local objectives of federated learning (FL) in dynamic heterogeneous systems, an adaptive federated meta learning framework is proposed. It inputs device conditions into a multimodal learning structure to optimize global FL and time estimation models, while allowing local model adaptation through meta-parameters. This framework enables heterogeneous devices to effectively contribute without hindering system-wide goals.

In collaborative perception, vehicles' unsynchronized clocks, mobility, and environmental dynamics can misalign locally extracted features, reducing fusion accuracy. We address these issues by designing clock synchronization and feature alignment, along with a feature-vehicle selection scheme via a hierarchical multi-armed bandit algorithm for latency requirements. Simulation results show the superiority of our approaches for improving the accuracy and efficiency of collaborative perception.