Non-myopic approaches to sensing and surveying
My name is George, and I am a 2nd Year PhD student and a member of the Distributed Algorithms group. I moved to Liverpool in 2017 to study for an integrated Masters in Aerospace Engineering.
Throughout my time at Liverpool, I have had an eclectic mix of jobs and internships, from a shelf stacker at Sainsbury's in Woolton, to running my own tutoring business. Liverpool has provided me with countless opportunities. I love the outdoors and to counter my normal day of sitting in front of a screen, I try to remain in balance and spend as much time as I can exploring the scenic parts of the local area.
Family is important to me; I was brought up by totally selfless parents who would give me everything they have, and I am forever grateful to them for this. I now have my own smaller scale family in Liverpool, living with my girlfriend and our newest addition, Orbit the border collie.
I have recently started a company with a good friend I met at university, who also studied aerospace engineering. We have recently secured funding from Innovate UK to work on a project to develop machine learning for the high value manufacturing sector.
My research
My research is based in algorithm development. Predominantly the non-myopic kind for tracking and sensor management type activities. Myopia is defined as the quality of being short-sighted, and algorithms that operate under this regime are considered greedy as they only choose what is of immediate benefit to them.
Imagine an autonomous ground vehicle, traversing the earth to keep track of a bird in the sky, which is about to fly over a small river. Under myopic planning the ground vehicle would follow the bird to the riverbank, realise it is now out of options and therefore lose track of the bird. Under non-myopic planning, the ground vehicle would foresee that it needs to head for the bridge adjacent to the bird’s flightpath to keep track of its target past the river.
I am taking a metric driven approach to solving this problem. Many existing implementations utilise information theory, but it is not exactly clear what you are explicitly trying to optimise for in this case.
The work I am doing is using a metric, known as the generalized optimal sub-pattern assignment (GOSPA) metric which can, and is, used to quantify the performance of a tracker as it penalises for the three quantities of interest in a target tracking scenario. The localisation error (how wrong our estimates are), the missed targets (if there is a target and we don’t think there is) and the false targets (if we think there is a target but there is not).
As the GOSPA metric can be used to effectively measure performance, it intuitively makes sense to build it into the decision-making part of the algorithm and use it as a driver to determine which actions are optimal.
What are your next steps?
I recently presented my work at the international symposium on information fusion conference in South Carolina, this was a great experience and allowed me to gain invaluable feedback from peer reviewers and expert audience members.
The aforementioned Innovate UK project is also now underway so I will be working on this alongside my studies and research.
My 3-minute thesis
The 3MT was a great experience. I learned a lot through the process, such as how to condense a complex proposition into less technical language and a short time frame. I met some great people throughout, including other students and the organisers, who put on a great event for the final.
If the opportunity were to present itself again, I would be more than happy to re-enter, being able to spend time refining my pitching skills and spending an afternoon listening to the other students. The support provided throughout and the feedback from the referees enabled me to reflect on the areas that need improvement. This is very helpful for my CPD as I now have some areas to focus on.