Reinforcement learning-based resource allocation for multi-task multi-codebook semantic communications

Abstract

Task-oriented semantic communications (SC) compress and transmit the meaning of data to meet end-task goals, e.g., perception or decision accuracy, using far fewer bits than classical source-channel pipelines. As these systems scale, two realities dominate deployments: (a) devices must simultaneously serve heterogeneous tasks such as reconstruction for telepresence, segmentation for autonomy, and classification for analytics, and (b) encoders increasingly rely on multi-codebook hierarchies that share low-level semantics across tasks while specializing higher layers. These requirements expose a new optimization problem: how to jointly control radio resources, including subcarriers, power, code rate, and per-task codebooks in real time to optimize semantic utility under power and latency constraints. We address this by developing a practical controller, named ReSCo-SAC, that integrates a unified hierarchical multi-codebook design for multi-task semantics and an entropy-regularized reinforcement learner based on Soft Actor-Critic (SAC). The hierarchical design supplies vertical cross-task structure and a top-down quick-search mechanism to accelerate codeword lookup and prune the codebook choices online, and the learner turns semantic feedback into resource decisions that track the marginal semantic gain. We formalize the system model, an operational objective, and a hybrid action space that combines continuous radio controls with discrete codebook selections. We then detail an implementation that projects relaxed actions to a feasible OFDMA schedule, couples codebook choice to the hierarchical search, and uses online semantic measurements to tune the controller. Extended simulation results demonstrate that the proposed multi-codebook structure and entropy-regularized ReSCo-SAC can perform robust, resource-efficient SC in multiple tasks.