A Goal-Oriented Effective Communication Scheme Via Action Innovations

In future wireless communication, an exponential increase in connected devices and the need for safety-critical and data-intensive applications like Autonomous Vehicles (AVs) and eXtended Reality (XR) will render large-scale intelligent agent deployments infeasible. To address this, we propose a novel framework where a centralized Smart Service Provider (SSP) transmits actions to agents, balancing effectiveness and transmission costs. Our approach leverages the observation that both semantic information extraction from raw sensor data and action generation from semantic states are many-to-one mappings. We model action decisions using stochastic policies from Reinforcement Learning (RL) literature. A key novelty is our focus on when to transmit. Namely, the SSP, observing what an agent observes, estimates the agent’s likely action. Based on the weighted statistical distance between the SSP’s policy and the agent’s estimated policy, the SSP decides whether an "action innovation" has occurred, and if so, transmits an action to the agent. Simulations demonstrate that transmitting only 10% of the time is sufficient to satisfy goals with up to 1% deviation, and compression rates up to 8,108:1 are achievable compared to existing standards. This "effective communication" scheme significantly reduces transmission overhead while maintaining high performance.