Clustering inference errors

An important aspect of artificial intelligence is being able to make inferences (guesses) about data. However, all systems make errors. This project involves developing new clustering algorithms suited to identifying errors that have a similar cause. As a case study, it will consider errors in an algorithm for identifying the amount of electricity that each household appliance consumes, based only on measurements of the entire household’s power consumption. This information is useful to households wishing to reduce their electricity consumption, go “off-grid” using solar panels and a batter, or evaluate a time-of-use plan offered by their retailer.

In particular, this data set contains many clusters in which errors appear to follow simple curved lines, when viewed in “feature space”. This is suited to an approach called manifold clustering, which attempts to choose suitable curves, balancing the need for the curves the be smooth with the need for them to fit the data closely.

Expected background: Artificial intelligence, geometry and basic calculus.

Preferred background: As above.

Supervisors Lachlan Andrew and Junhao Gan