Poster Abstract: Always Heading for the Peak: Learning to Route with Domain Knowledge

Routing optimization is a well-known yet complicated research topic for decades. On one hand, routing algorithms usually have complex inputs (e.g., structural topology), outputs (e.g., sequential paths for different demands), and optimization goals (e.g., link utilization). On the other hand, due to the rapid change of demands, routing algorithms must be fast enough to make online decisions. Existing methods usually make a tradeoff between performance and efficiency. Offline algorithms [1] tend to cast the problem into a Linear Program (LP), offering near-optimal solutions yet taking up to tens of seconds, which are too heavyweight for rapidly changing networked systems. Online algorithms, such as shortest path (SP) and weighted shortest path (WSP), are broadly used, but their performance is far from optimal due to the ignorance of network status. Conventional heuristic algorithms fail to efficiently capture the dynamics inside varying network demands, which compromises the performance, as shown in Table I.