Multiagent Meta-Reinforcement Learning for Adaptive Multipath Routing Optimization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Network Engineering

Background:

• Packet network routing is a complex challenge in distributed systems.
• Existing routing optimization policies struggle with dynamic network conditions.
• Multiagent reinforcement learning (RL) offers a promising approach for autonomous network management.

Purpose of the Study:

• To develop advanced multiagent reinforcement learning algorithms for efficient packet network routing.
• To address the challenges of multitask learning in network routing due to interdependent node policies and traffic demands.
• To optimize network performance under both fixed and time-varying traffic conditions.

Main Methods:

• Modeling the routing problem as a networked multiagent partially observable Markov decision process (MDP).
• Proposing two novel model-free multiagent RL algorithms: multiagent proximal policy optimization (MAPPO) and multiagent metaproximal policy optimization (meta-MAPPO).
• Designing a distributed implementation framework for MAPPO leveraging exploration-exploitation separability.

Main Results:

• MAPPO and meta-MAPPO demonstrate excellent performance in optimizing network routing.
• The proposed algorithms effectively handle both fixed and time-varying network traffic demands.
• Simulation results show superior performance compared to existing routing optimization policies.

Conclusions:

• The developed MAPPO and meta-MAPPO algorithms represent a significant advancement in multiagent RL for network routing.
• The proposed methods offer robust and efficient solutions for optimizing packet network performance.
• This research provides a practical framework for implementing advanced RL-based routing in real-world networks.