Learning Multi-Access Point Coordination in Agentic AI Wi-Fi with Large Language Models

🕒 Published (v1): 2025-11-25 06:29 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper proposes an agentic AI framework for Wi-Fi multi-access point coordination (MAPC) where each access point is an autonomous LLM agent that negotiates adaptive transmission schedules via natural language dialogue. Each agent is equipped with a modular cognitive architecture—short-term sliding-window memory, long-term RAG-based knowledge base, tool use, and chain-of-thought prompting—to learn Co-TDMA vs. Co-SR tradeoffs at runtime. DeepSeek-R1- and GPT-4o-powered agents outperform the optimized Wi-Fi 6 spatial reuse baseline by up to 85% normalized throughput in Co-SR-favored scenarios.

Problem

Conventional MAPC protocols for dense Wi-Fi deployments (IEEE 802.11bn / Wi-Fi 8) rely on static, protocol-defined heuristics that cannot adapt to dynamic interference conditions, varying topologies, or the Co-TDMA/Co-SR tradeoff. Prior AI-based approaches (multi-armed bandits, HMARL) use task-specific small models with poor generalization to unseen environments and narrow coordination modes.

Method

Each AP is modeled as an LLM agent with a structured cognitive architecture running within a trigger-frame TXOP:

1. Coordination protocol: A sharing AP polls neighboring shared APs, forming a group. Over \(T\) negotiation rounds, each round of \(L\) time slots proceeds as: proposal broadcast by sharing AP → autonomous schedule selection by shared APs → data transmission → tool-invoked outcome retrieval → feedback loop.

2. LLM brain + CoT workflow: Three-phase chain-of-thought per round—Evaluation (score last round using a reward function: +reward for successful co-transmission, penalty for collision or idle slot), Reflection (synthesize outcomes into strategic insight, update long-term memory if performance exceeds threshold), Action Generation (emit transmission schedule \(a_k \in \{0,1\}^L\) and natural language coordination message).

3. Hybrid memory: Short-term memory is a sliding window of \(W=5\) recent rounds injected into the contextual prompt. Long-term memory is a RAG knowledge base of up to 10 curated exemplars; retrieval uses cosine similarity over embeddings; updates replace the lowest-scoring exemplar in a similar cluster only when the new round outperforms it.

4. Prompt engine: Two-part prompt—static core prompt (identity, scoring rules, ICL examples, format) plus a dynamic contextual prompt (inter-agent messages, short-term history, RAG-retrieved exemplars).

5. Tool use: get_transmission_outcome() interfaces with the network monitor and translates raw channel feedback into structured text for LLM consumption.

Multi-agent dialogue is implemented via Microsoft AutoGen.

Key Contributions

• First multi-LLM-agent framework for MAPC, enabling distributed, dialogue-based coordination without fine-tuning.
• Hybrid memory design (short-term sliding window + performance-gated RAG) enabling both tactical adaptation and strategic self-improvement.
• Demonstrated heterogeneous LLM interoperability (GPT-4o ↔ DeepSeek-R1 agents in the same network).
• Backward compatibility: agentic APs coexist with legacy CSMA/CA APs without disruption.
• Ablation isolating contributions of reflection, inter-agent negotiation, short-term memory, and long-term memory.

Results

All throughput values are normalized (fraction of total time used for successful transmissions); baseline is optimized Wi-Fi 6 OBSS/PD spatial reuse (AC_BE traffic, exhaustive parameter search):

• 2AP Co-SR-Favored: GPT-4o: 1.85×, DeepSeek-R1: 1.87× vs. baseline 1.05×
• 3AP Co-SR-Favored: GPT-4o: 2.06×, DeepSeek-R1: 1.80× vs. baseline 1.35×
• 2AP Co-TDMA-Favored: GPT-4o: 1.00×, DeepSeek-R1: 0.91× vs. baseline 0.84×
• 3AP Co-TDMA-Favored: GPT-4o: 0.91×, DeepSeek-R1: 0.90× vs. baseline 0.86×
• Ablation (2AP Co-TDMA, normalized to full system = 1.00):
• w/o Reflection: 1.00 (no degradation in TDMA; loses Co-SR exploration)
• w/o Inter-Agent Negotiation: 0.40 (catastrophic in TDMA), 1.70 in Co-SR
• w/o Short-Term Memory: 0.71 TDMA, 1.81 Co-SR
• w/o Long-Term Memory: 0.66 TDMA, 1.75 Co-SR
• Coexistence test: LLM-APs achieve higher throughput than legacy-only CSMA/CA in mixed networks (3–6 OBSS nodes) while not degrading legacy AP throughput.

Limitations

• Evaluated only in simulation; real-world Wi-Fi latency and packet overhead for LLM inference are not modeled.
• LLM inference latency is not compared against TXOP timing constraints (400 µs rounds)—feasibility of on-device or edge inference is unaddressed.
• Knowledge base capped at 10 exemplars; scalability of RAG to large, heterogeneous topologies is untested.
• Ablation shows near-zero gain in Co-TDMA-favored scenarios over the baseline, leaving headroom for reflection-free simpler policies in interference-heavy deployments.
• Multi-vendor heterogeneity tested only for 2-AP GPT-4o + DeepSeek-R1; larger mixed fleets not evaluated.

Relevance to Harnesses / Meta-Harnesses

This paper is a concrete instantiation of an agentic workflow harness applied to a real-time distributed control problem: the cognitive architecture (static core prompt + dynamic contextual prompt + RAG retrieval + tool invocation + CoT reflection loop) is essentially a per-agent mini-harness that structures each decision cycle. The two-part prompt design—separating durable agent identity from ephemeral context injection—directly mirrors the system-prompt / dynamic-context split common in meta-harness frameworks. The use of AutoGen as the multi-agent orchestration layer further situates this as a domain-applied harness study, demonstrating that harness patterns (memory management, tool use, reflection loops, structured output formats) transfer to non-NLP domains like wireless scheduling. The ablation results quantify the contribution of individual harness components (memory, negotiation, reflection), providing rare empirical evidence about which sub-harness modules matter most in a deployed agentic loop.