AI-Adaptive Network Routing

Routing Mode

AI-Adaptive Routing

Traffic

Packet spawn rate 3.0/s

AI learning rate α 0.30

Sim speed 1.0×

Live Metrics

Avg latency

— ms

Throughput

0.0/s

Delivered

Dropped

How it works

Each link in the mesh has a base cost (proportional to its length) plus a congestion penalty that grows with how many packets are travelling on it right now, plus an occasional injected spike. Routing decides, for every new packet, the cheapest path from source to destination using Dijkstra's algorithm over an edge-cost table:

cost(edge) = base · (1 + 0.35·concurrency) · spikeFactor
path = Dijkstra(graph, cost, source, destination)

The two modes differ only in what cost table Dijkstra sees:

Static shortest-path — always uses the fixed base cost (physical distance only), exactly like classic distance-vector/link-state routing with static weights. It cannot see congestion, so it keeps funnelling packets into a spiked or crowded link.
AI-adaptive — maintains a learned cost estimate per edge, updated after every packet finishes crossing it with an exponential moving average:

estimate ← (1 − α)·estimate + α·(observed latency)

A higher learning rate α reacts to congestion faster but is noisier; a lower α is smoother but slower to notice a spike. Packets entering an already-busy edge face a rising drop probability, modelling buffer overflow at a real router. Trigger a congestion spike and compare: AI-adaptive routing detects the slow edge within a few packets and reroutes around it; static routing keeps sending traffic straight into it until the spike ends.

Packet in flight Link — low load Link — congested Link — spiked/overloaded

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.