The airport stand assignment problem (SAP) is concerned with allocating arriving and departing flights to parking areas ("stands") in order to maximize the airport's operational efficiency. This summer, Petuum was approached by a client to address the SAP for their airport's needs.Â
I was the first engineer in the company to work on this project, so I had the exciting opportunity to lay much of the technical groundwork for tackling the SAP. First, I mathematically formulated the SAP: initially as an mixed integer program, then as a partially observable Markov decision process. This allowed me to articulate exactly what data we required from our client. Then, in Python, I implemented a deep reinforcement learning agent end-to-end to begin addressing the SAP, which included me designing neural networks for policy function approximation and implementing RL training processes (e.g. REINFORCE, A2C) in Tensorflow.
Throughout the project, I conducted meetings with a 17-person team of product managers, software engineers, and AI researchers to articulate proposed solutions to the stand assignment problem.
On the side, I created a user journey and initial wire-framing for a client’s mobile fashion app, and analyzed competing services for an HR bot developed by Petuum, Kaibot.
