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

TL;DR

The paper tackles the need for Adaptive Intent Management Functions (IMFs) that can handle changing customer intents and priorities without retraining. It introduces a supervisor policy $\pi_g$ that coordinates multiple MARL subsystems and augments it with a DUN-based feature engineering module to encode the utility context, enabling runtime adaptation to different utility function forms such as linear, logarithmic, and quadratic. The approach preserves a fixed training regime while allowing execution-time changes in the function form and priorities, demonstrated via a GlobeCom network emulator across diverse scenarios, including extended action spaces like Move-UE-context and Auto-Scale. Results indicate improved generalization (lower Integral Absolute Error) and faster convergence compared with baselines, suggesting significant practical impact for live, resource-constrained 6G networks.

Abstract

Intent Management Function (IMF) is an integral part of future-generation networks. In recent years, there has been some work on AI-based IMFs that can handle conflicting intents and prioritize the global objective based on apriori definition of the utility function and accorded priorities for competing intents. Some of the earlier works use Multi-Agent Reinforcement Learning (MARL) techniques with AdHoc Teaming (AHT) approaches for efficient conflict handling in IMF. However, the success of such frameworks in real-life scenarios requires them to be flexible to business situations. The intent priorities can change and the utility function, which measures the extent of intent fulfilment, may also vary in definition. This paper proposes a novel mechanism whereby the IMF can generalize to different forms of utility functions and change of intent priorities at run-time without additional training. Such generalization ability, without additional training requirements, would help to deploy IMF in live networks where customer intents and priorities change frequently. Results on the network emulator demonstrate the efficacy of the approach, scalability for new intents, outperforming existing techniques that require additional training to achieve the same degree of flexibility thereby saving cost, and increasing efficiency and adaptability.

Figures (8)

• Figure 1: Block diagram depicting performance of baseline method (Fig 1a) and thereby need for improved convergence for unseen situations i.e when intent priority or function form changes. The plot in 1b depicts the significant improvement in convergence for the same unseen situations i.e generalization abilities when our proposed method is leveraged.
• Figure 2: Block diagram of the proposed methodology. The supervisor agent assigns sub-goals to the lower-level agents by observing current and past performance, global goals (derived from intents) and utility context $n_{t+1}$. The observed KPI values ($x_i$ for agent $i$) for all agents are send to feature engineering block (as per \ref{['eqn:feature']}) and extracted features($y_i$) to DUN to create utility context $n_{t+1}$. The four types of lower-level agents used in this work are: The priority MARL systems of 3 agents which controls unique packet priorities for CV, URLLC and mIoT. The MBR MARL system which changes Max Bit rate. The third corresponds to Move-UE-context system which contains two agents (one for URLLC and another for mIoT). The fourth one is a single RL agent. The areas with semi-transparent(gray) boxes are our contributions when compared to existing literature.
• Figure 3: Execution results of supervisor agent trained using baseline method (a,b) and proposed method (d,e) on scenarios 1 and 2. Both methods perform equally well in execution phase as long as priority and utility is not changed. Plot (c) corresponds to testing the baseline model trained in scenario 1 on scenario 2. Plot (f) corresponds to proposed method where baseline model trained in scenario 1 is tested on scenario 2. From the plots, it is evident that existing baseline approach fails to generalize well whereas the proposed approach far outperforms the existing with change in priority and utility function.
• Figure 4: IAE values of all the KPIs vs $P_i$ for CV with $P$ at $1$ for URLLC and mIoT. No change in function form.
• Figure 5: Evaluation on Form2 \ref{['eq:log']} of utility function: Variation in IAE values of all KPIs vs $P_i$ values for prioritized intent keeping $P_i$ for other intents fixed at $1$. Solid lines correspond to proposed approach whereas dashed lines are the baseline.