Lab2 PSC, 11:00-11:30
Dynamic Control of a Queueing System with Abandonments by Terry James, STOR-i PhD student
Defensive surveillance is of great importance in the modern world, motivated by the threats faced on a daily basis and the technology which now exists to mitigate these threats. Our research project considers a defensive surveillance problem of a surveillance resource responsible for a number of public areas, each of which is a potential target for an adversary. The problem is thought of as a multiclass queueing system with customer abandonments. The resource or server must identify suitable surveillance policies in the face of strategic adversaries wishing to remain covert and avoid detection in amongst civilians.
This leads to the question of what is meant by a policy and how does this impact the system? To address this question we consider a strongly related queueing control problem which also helps to form the basis of the surveillance problem. The queueing control problem pits a single server against a multi-class system in which customers carry rewards, but can also abandon the system taking these rewards with them as they renege. How then should the server act to maximise his average reward earned from the system?
The forum talk will consider approaches to constructing the heuristic policies required in high-dimensional systems due to dimensionality problems. One of our main approaches is based on the concept of one-step policy improvement. Here the main challenge is the approximation of bias functions. Various approximation methods have been explored in our research and are the subject of ongoing thinking.