Intelligent resource allocation for knowledge-driven semantic communication networks

The future communication system is undergoing a paradigm shift from "transmitting bits" to "conveying meaning," with semantic communication (SemCom) emerging as its core technology. SemCom’s ultra-high efficiency relies on a shared knowledge base (KB) between sender and receiver. However, in dynamic environments, the KB inevitably becomes outdated, causing a sharp decline in semantic efficiency. To restore system performance, the KB must be updated and synchronize. The core insight of this thesis is that the maintenance process of the KB and the transmission of semantic data itself compete for the same limited wireless resources. This creates a novel and fundamental dynamic trade-off between the immediate utility of transmission and the long-term benefits of KB updates. Existing literature often overlooks this coupling effect, failing to balance the immediate utility of data transmission against the long-term reliability of KB maintenance. First, to address the KB staleness trade-off, this thesis models knowledge obsolescence as a quantifiable state variable. Based on this, we formulate the dynamic tradeoff between semantic transfer utility and knowledge consensus cost as a 0-1 Mixed-Integer Non-Linear Programming (MINLP) problem. We propose an online scheduling algorithm based on model predictive control (MPC) and iterative marginal cost allocation (IMCA) to efficiently solve this tradeoff. Second, the aforementioned resource allocation problem, along with other wireless networking applications, mathematically manifests as an NP-hard 0-1 MINLP. Traditional optimization solvers struggle to scale due to exponential complexity, while pure reinforcement learning (RL) methods suffer from inefficient search due to blind exploration. Thus, we aim to address the aforementioned challenges by constructing a unified resource optimization framework. Specifically, the main research contribution is proposing a unified 0-1 mixed optimization framework, which models discrete decision processes as Markov Decision Processes (MDPs). Its core innovation lies in solving the continuous relaxation of the original problem as guidance and theoretically proving that the neighborhood of this relaxed solution defines a high potential zone (HPZ), thereby transforming the RL agent’s exploration from blind trial-and-error to efficient guided search. Third, regarding the KB synchronization mechanism itself, existing consensus protocols are designed for wired networks with deterministic fault models, making them overly conservative and inefficient for probabilistic wireless environments. To bridge this gap, this thesis builds an availability-robustness analysis framework for consensus protocols, for consensus protocols used in tasks like KB maintenance, this paper introduces an innovative dual-metric reliability model. This model quantifies the inherent tradeoff between availability and robustness, guiding the optimal design of the quorum through solving a constrained optimization problem. Finally, this thesis designs and builds a multi-user SemCom physical platform based on non-orthogonal multiple access (NOMA). We define semantic throughput (STU) as the optimization objective and propose an improved watering algorithm to address the non-convex semantic-aware power allocation problem. Experimental results validate the significant performance gains of the proposed approach at the semantic level.