Multi-agent adaptive routing by multi-head-attention-based twin agents using reinforcement learning

A regular condition, typical for packet routing, the problem of cargo transportation, and flow control, is variability graph. Reinforcement learning based adaptive routing algorithms are designed to solve with this condition. However, significant changes in graph, existing require complete retraining. To handle challenge, we propose a novel method on multi-agent modeling twin-agents which new neural network architecture multi-headed internal attention proposed, pre-trained within framework multi-view paradigm. An agent such paradigm uses vertex as an input, twins main placed at vertices graph select neighbor object should be transferred. We carried out comparative analysis DQN-LE-routing algorithm two stages: pre-training simulation. In both cases, launches were considered by changing topology during testing or Experiments have shown that proposed adaptability enhancement provides global increasing delivery time only 14.5 % after occur. The can used problems complex path evaluation functions dynamically topologies, example, transport logistics managing conveyor belts production.