MAMO: Multi-Agent System for Constrained Multi-Objective Optimization

Many real-world decision-making problems in computing and networking systems involve minimizing costs while adhering to performance constraints. In dynamic settings, reinforcement learning (RL) is often employed to solve these problems by combining costs and constraint violations into a single reward signal, typically using weighted penalty terms. However, the effectiveness of the learned policy heavily depends on the manual selection of these weights, which can be challenging, especially in environments where the relative importance of objectives and constraints changes over time. To overcome this limitation, researchers have introduced MAMO (Multi-Agent system for Multi-Objective constrained optimization). MAMO takes a novel approach by formulating the selection of these crucial reward weights as a separate learning problem, effectively decoupling the core task execution from the design of the objective function itself. This is achieved through a multi-agent RL framework. By allowing the system to learn the optimal trade-off between optimizing the primary objective and satisfying constraints autonomously, MAMO represents a significant step towards more robust and self-adapting RL-based solutions. This approach promises to simplify the deployment of RL in complex, dynamic environments where manual tuning of reward weights is impractical or suboptimal.