Title: KISS: Keeping it Simple and Slotted when Learning to Communicate over Wireless

Abstract

Achieving efficient and equitable random channel access remains a persistent hurdle in distributed wireless systems. While current approaches often tackle specific limitations regarding timing, periodicity, or central control, they predominantly depend on static heuristics. Inspired by recent progress in machine learning (ML), this study explores the potential of ML agents to autonomously develop access strategies that are both fair and efficient, and whether such learning processes can provide fresh perspectives on Medium Access Control (MAC) architecture. Rather than introducing a ready-to-deploy protocol, our objective is to determine if decentralized learning can approximate or rediscover theoretically optimal random-access mechanisms under minimal assumptions. To achieve this, we utilize an off-policy Double Deep Q-Network (DDQN) integrated with Bayesian inference to train agents within a slotted channel environment. The proposed method is entirely distributed, involving independent multi-agent learners, and operates fully online without the need for pre-training. It is stochastic rather than periodic, functioning without any coordination or explicit communication between nodes. Comprehensive simulations demonstrate that the learned strategy adjusts to fluctuating network conditions, attaining near-theoretical efficiency while preserving fairness. Furthermore, ablation studies indicate that the agent’s behavior mirrors slotted ALOHA, characterized by a dynamically tuned transmission probability. This similarity prompts us to name the method KISS: Keeping It Simple and Slotted.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC