Agentic AI for Intent-driven Optimization in Cell-free O-RAN

1. Problem Statement

The paper addresses the challenge of managing Cell-free Open Radio Access Networks (O-RAN) using autonomous, intent-driven optimization. While O-RAN architecture (disaggregated into O-RUs, O-DUs, and O-CUs) enables flexible control, current AI-driven approaches face two main limitations:

• Complexity of Intents: Existing works typically handle simple, independent intents. They struggle with complex intents that require coordination among multiple agents (e.g., simultaneously minimizing energy consumption while guaranteeing specific minimum data rates for different users).
• Scalability of LLM Agents: Deploying separate Large Language Models (LLMs) for every agent in the near-real-time (near-RT) RIC creates prohibitive memory overhead, limiting the scalability of autonomous RANs.

The core optimization problem involves maximizing a utility function (either aggregate data rate or energy savings) subject to:

1. Minimum Rate Constraints: $r_k \geq R^{\min}_k$ for all users.
2. Power Constraints: Total transmit power per active O-RU must not exceed $P^{\max}$.
3. Binary Activation: Deciding which O-RUs to activate ($z_l \in \{0, 1\}$) to save energy.

This is a non-convex, mixed-integer programming problem (NP-hard in energy-saving mode).

2. Methodology: Agentic AI Framework

The authors propose a multi-agent framework operating across different timescales within the O-RAN architecture (Non-RT RIC and Near-RT RIC). The framework utilizes Large Language Models (LLMs) for reasoning and intent translation, combined with Deep Reinforcement Learning (DRL) and optimization algorithms.

A. Agent Roles and Workflow

1. Supervisor Agent (Non-RT RIC):

   • Acts as an rApp.
   • Receives natural language operator intents (e.g., "Enter energy-saving mode but guarantee 50 Mbps for User 3").
   • Translates these intents into mathematical objectives (utility functions) and specific constraints ($R^{\min}_k$).
   • Forwards these to the Near-RT agents via the A1 interface.

2. User Weighting Agent (Near-RT RIC):

   • Acts as an xApp.
   • Determines user priority weights ($\alpha_k$) for precoding.
   • Uses a Memory Module to retrieve relevant prior experience (historical coefficients) based on current channel conditions ($\beta_{k,l}$) and rate requirements.
   • Updates Lagrange multipliers ($\mu_k$) to satisfy rate constraints.

3. O-RU Management Agent (Near-RT RIC):

   • Acts as an xApp.
   • In energy-saving mode, determines the set of active O-RUs ($z_l$).
   • Utilizes a Multi-Agent Proximal Policy Optimization (MAPPO) algorithm. Each O-RU is an agent that observes local fading and rate violations to decide its activation state.
   • The reward function penalizes active O-RUs, rate violations, and frequent state switching.

4. Monitoring Agent (Near-RT RIC):

   • Continuously measures user data rates.
   • Coordinates the User Weighting and O-RU Management agents.
   • If minimum rates are violated, it triggers feedback loops: increasing user priority weights ($\alpha_k$) or violation penalty coefficients ($\lambda_k$) to reactivate O-RUs.

B. Optimization Algorithms

• Precoding: Uses a Weighted Minimum Mean Square Error (WMMSE) approach to iteratively update precoding matrices ($V_{k,l}$) based on the calculated weights.
• Energy Saving (DRL): The O-RU activation problem is solved via MAPPO. Agents learn a policy to minimize active O-RUs while satisfying rate constraints through a shared critic network.
• Retrieval-Augmented Tuning: To accelerate convergence, the system stores key-value pairs (environment features $\to$ optimal coefficients) in a memory module. When a new intent arrives, it retrieves similar past experiences to initialize coefficients, skipping the initial search phase.

C. Scalability via QLoRA

To address the memory bottleneck of deploying multiple LLMs:

• A single shared LLM is deployed in the Near-RT RIC.
• Quantized Low-Rank Adaptation (QLoRA) is used to fine-tune this shared model for each specific agent task.
• The backbone weights are quantized to 4-bit (FP4) and frozen, while small, trainable low-rank adapters ($A$ and $B$ matrices) are trained for each agent. This preserves specialization while drastically reducing memory footprint.

3. Key Contributions

1. Intent-Driven Coordination: A novel framework where LLM-based agents collaborate to translate complex, natural language intents into coordinated control actions (precoding weights and O-RU activation) across different timescales.
2. Hybrid AI Architecture: Integration of LLMs for high-level reasoning/intent translation with DRL (MAPPO) for distributed decision-making and WMMSE for physical layer optimization.
3. Memory-Efficient Deployment: Demonstration of using a shared LLM with QLoRA adapters, enabling the deployment of specialized agents in resource-constrained Near-RT RICs without the overhead of separate full-precision models.
4. Retrieval-Augmented Convergence: A mechanism to store and retrieve optimal control coefficients based on environmental similarity, significantly speeding up the adaptation to new intents.

4. Performance Results

Simulations were conducted in a cell-free O-RAN with 50 O-RUs and varying numbers of users.

• Energy Efficiency:

  • In energy-saving mode, the proposed framework reduced the number of active O-RUs by 41.93% compared to three baselines:
    1. Full-power mode (all O-RUs active).
    2. Greedy algorithm (activating O-RUs based on strongest channel).
    3. DRL with Gradient Ascent (DRL+GA) without agent coordination.
  • The DRL+GA baseline performed poorly due to a lack of coordination, causing instability where penalty coefficients grew too fast, forcing unnecessary O-RU activations.

• Memory Efficiency:

  • Using the QLoRA approach with a shared 7B parameter model, the framework reduced memory usage by 92% compared to deploying three separate full-precision (FP16) LLMs.
  • For a 7B model, memory dropped from ~45.7 GB (3 separate models) to 3.8 GB (1 shared + adapters).

• Intent Adaptation:

  • The system successfully switched between "Energy Saving" and "Utility Maximization" intents.
  • When switching intents, the retrieval-augmented memory allowed the system to skip the initial search phase, immediately applying previously learned optimal coefficients for similar scenarios.

5. Significance

This paper represents a significant step toward autonomous 6G networks. It moves beyond simple, isolated AI tasks to a collaborative agentic system capable of understanding high-level human intent and executing complex, multi-objective optimizations.

• Practicality: By solving the memory scalability issue via QLoRA, the paper makes the deployment of LLM-based agents in real-time network controllers (Near-RT RIC) technically feasible.
• Robustness: The coordination between the monitoring, weighting, and management agents ensures that aggressive energy-saving measures do not compromise Quality of Service (QoS), a critical requirement for commercial deployment.
• Future Direction: The work lays the groundwork for fully autonomous RANs where operators can manage network resources via natural language, with the AI handling the underlying complexity of resource allocation and energy management.