Non-parametric supervised learning

Lead Research Organisation: University of Bristol
Department Name: Mathematics

Abstract

In my research I will explore a new approach to the Contextual Multi-Armed Ban-dits problem (CMAB) [1], which is based on a novel method for nonparametric regression. CMAB is a sequential decision making problem in which one aims to maximise some "re-ward" over a given time period. As the data is observed, the proposed method involves partitioning the data into disjoint subsets, for which simple models are sufficient. Moving forward I will assess the trade off between predictive performance and computational cost of the partitioned approach. Additionally, my research will investigate how partitioning schemes can be tailored to the CMAB setting. A well tuned partitioning scheme has the potential to isolate situations where the decision maker has little experience, leading to a more accurate quantification of uncertainty. Developing the general methodology associated with data adaptive partitions, and how best to construct them, is a key focus of my work. As the dimension of the data grows, the set off possible partitions becomes larger, and thus more expensive to search. It is therefore important to establish scalable methods for proposing suitable partitions and recognising when to terminate the search. In the current approach this is achieved by placing a Bayesian hierarchical model across the partitioned data. This allows information to be shared between subsets, which can be used to help identify whether a given subset requires further partitioning.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S023569/1 01/04/2019 30/09/2027
2266135 Studentship EP/S023569/1 23/09/2019 22/09/2023 Douglas Corbin