Reinforcement-Learning-Optimized Multi-Agent Intelligence for Semantic Feature Selection in AI-Native 6G Wireless Indoor Localization Networks

Abstract

Abstract—The transition from fifth-generation (5G) to sixth-generation (6G) wireless systems demands networks that are not only faster but cognitively autonomous, capable of sensing, reasoning, and reconfiguring themselves without continuous human supervision. This paper proposes a Multi-Agent Intelligence (MAI) framework for AI-native 6G wireless indoor localization that unifies five cooperating agents, a Spectrum Agent for semantic feature ranking, a Security Agent for anomaly and intrusion screening, a Reconfigurable Intelligent Surface (RIS) Agent for channel enhancement, a Drone Agent for coverage assistance, and a federation of Base Station Agents that perform privacy-preserving federated learning — orchestrated by a tabular Q-learning Reinforcement Learning (RL) Coordinator that adapts the number of transmitted semantic features to instantaneous network load, latency sensitivity, and security posture. The framework is evaluated on a wireless indoor localization dataset comprising seven Wi-Fi access-point received-signal-strength-indicator (RSSI) features mapped to four room-level location classes. A centralized Random Forest baseline attains 99.95% (± 0.10%) five-fold cross-validated accuracy, establishing an upper bound on the achievable task performance. The fully federated configuration trained over six Base Station Agents across thirty aggregation rounds attains 99.40% test accuracy and a weighted F1-score of 0.9940 while never exposing raw client data. The proposed RL-optimized multi-agent configuration, which selects four of seven semantic features under the learned policy, attains 99.80% test accuracy, a weighted F1-score of 0.9980, and a 42.86% reduction in transmitted feature dimensionality relative to full-feature transmission, demonstrating that cooperative, agentic, and privacy-preserving wireless intelligence can match centralized performance while substantially reducing communication overhead. Mutual-information-based semantic ranking and reward-shaped reinforcement learning are formalized mathematically and justified individually, while security screening, channel/coverage modeling, and federated aggregation are described procedurally. Comparative analysis against recent state-of-the-art federated, multi-agent, and semantic-communication approaches confirms that the proposed framework achieves a favorable accuracy-versus-communication-cost operating point. The results support the broader IEEE Wireless Communications argument that future 6G radio access networks should evolve from centralized, monolithic optimization toward distributed, agentic, self-optimizing, and privacy-aware wireless intelligence.