Faculty Sponsor: Victoria Manfredi
Nathan Wu
Nathan Wu is a rising junior (’28) from Plano, Texas. Nathan attended Plano West Highschool, and, outside of his studies, he has two dogs that he really loves named Pebble and Kiwi. Nathan is majoring in Mathematics, Computer Science, and Economics. He hopes to get into quantative trading.
Abstract: For some communication networks, like mobile wireless networks, the network conditions are highly dynamic due to device mobility, wireless link conditions, and varying amounts of traffic over time. These dynamic conditions, however, make communication using such networks difficult. For instance, packet-level decisions (such as choosing the best next hop for a packet) and device-level decisions (such as choosing which packet in the device’s queue to forward first) are difficult when a device’s neighborhood is continually changing.
In this work, we focus on device-level decisions, given a packet-level decision agent. We specifically design a device-level reinforcement learning agent to compare queued packets using state and action features such as packet q-value, time spent at device, remaining packet lifetime, and next-hop queue length. These features are input into a neural network which outputs a q-value describing the goodness of selecting each candidate packet. A reward function is then developed to balance reliability, speed, fairness, and congestion, so that the device agent learns to choose packets that support overall network performance rather than simply following a fixed queue scheduling strategy such as FIFO.
Simulation data is converted into training examples by matching packet-layer outcomes back to device-layer queue scheduling decisions. The resulting model can then score future packet choices and support adaptive scheduling in dynamic communication networks. Final evaluation will compare the learned scheduler against baseline queue policies using metrics such as delivery rate, delay, forwarding cost, and queue drops.
nwu_project26