Resilient Packet Forwarding: A Reinforcement Learning Approach to Routing in Gaussian Interconnected Networks with Clustered Faults

TL;DR

This paper introduces a reinforcement learning-based routing method for Gaussian Interconnected Networks that adaptively bypasses faults, significantly improving packet delivery rates under high fault densities compared to traditional greedy routing.

Contribution

It presents a novel RL routing scheme using PPO tailored for fault-prone Gaussian networks, outperforming greedy algorithms in reliability and congestion management.

Findings

RL routing achieves 0.95 PDR at 40% fault density

Outperforms greedy routing in delivery rate and fault resilience

Effective under low network load with higher delivery rates

Abstract

As Network-on-Chip (NoC) and Wireless Sensor Network architectures continue to scale, the topology of the underlying network becomes a critical factor in performance. Gaussian Interconnected Networks based on the arithmetic of Gaussian integers, offer attractive properties regarding diameter and symmetry. Despite their attractive theoretical properties, adaptive routing techniques in these networks are vulnerable to node and link faults, leading to rapid degradation in communication reliability. Node failures (particularly those following Gaussian distributions, such as thermal hotspots or physical damage clusters) pose severe challenges to traditional deterministic routing. This paper proposes a fault-aware Reinforcement Learning (RL) routing scheme tailored for Gaussian Interconnected Networks. By utilizing a PPO (Proximal Policy Optimization) agent with a specific reward structure designed to penalize fault proximity, the system dynamically learns to bypass faulty regions. We compare our proposed RL-based routing protocol against a greedy adaptive shortest-path routing algorithm. Experimental results demonstrate that the RL agent significantly outperforms the adaptive routing sustaining a Packet Delivery Ratio (PDR) of 0.95 at 40% fault density compared to 0.66 for the greedy. Furthermore, the RL approach exhibits effective delivery rates compared to the greedy adaptive routing, particularly under low network load of 20% at 0.57 vs. 0.43, showing greater proficiency in managing congestion, validating its efficacy in stochastic, fault-prone topologies