# Towards Adaptive IMFs -- Generalization of utility functions in Multi-Agent Frameworks

**Authors:** Kaushik Dey, Satheesh K. Perepu, Abir Das, Pallab Dasgupta  
**arXiv:** [2405.07621](https://arxiv.org/abs/2405.07621)  
**Date:** 2024-05-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2405.07621v2/

## Key points

- The Intent Management Function (IMF) is crucial for future-generation networks
- Recent advancements focus on AI-based IMFs that can handle conflicting intents and prioritize global objectives
- Previous works utilized Multi-Agent Reinforcement Learning (MARL) techniques with AdHoc Teaming (AHT) approaches for conflict resolution in IMF
- This paper introduces a novel mechanism enabling IMF to generalize to different utility functions and adjust intent priorities at runtime without additional training
- Experimental results demonstrate the effectiveness of this approach, showcasing scalability for new intents and outperforming existing techniques
- Actions like Move-UE-context and Auto-Scale Limits are introduced to manage latency and power consumption efficiently in services like URLLC and mIoT
- Different types of lower-level agents are employed, enhancing the framework for flexible intent management in dynamic network environments

## Summary

, , , , 

The Intent Management Function (IMF) is a crucial component of future-generation networks, with recent advancements focusing on AI-based IMFs that can handle conflicting intents and prioritize global objectives based on predefined utility functions and priorities. Previous works have utilized Multi-Agent Reinforcement Learning (MARL) techniques with AdHoc Teaming (AHT) approaches for efficient conflict resolution in IMF. However, the adaptability of such frameworks to evolving business scenarios is essential for real-life deployment. This paper introduces a novel mechanism that enables IMF to generalize to different forms of utility functions and adjust intent priorities at runtime without requiring additional training. This flexibility is vital in dynamic network environments where customer intents and priorities frequently change. Experimental results on a network emulator demonstrate the effectiveness of this approach, showcasing scalability for new intents and outperforming existing techniques that necessitate additional training for similar levels of flexibility. In Experiment 5, the scalability of the proposed method is tested by increasing instances and action spaces within URLLC and mIoT services. The introduction of actions like Move-UE-context, which optimizes latency by shifting serving instances between nodes, and Auto-Scale Limits, which control the number of pods available for scaling applications, allows agents to manage latency and power consumption efficiently. Power optimization becomes crucial for mIoT services due to their large device count, with considerations given to energy costs at central versus edge sites. The proposed methodology includes a supervisor agent assigning sub-goals to lower-level agents based on performance observations, global goals derived from intents, and utility context updates. Different types of lower-level agents are employed, including priority MARL systems for unique packet priorities, MBR MARL systems for adjusting Max Bit rate, Move-UE-context systems for latency optimization, and single RL agents. These contributions enhance the existing literature by providing a comprehensive framework for flexible intent management in dynamic network environments. Overall, this paper highlights the innovative approach taken in enabling IMF generalization without additional training requirements while emphasizing its practical applicability in addressing evolving customer needs within modern networks.