The University of Southampton
University of Southampton Institutional Repository

Accounting for real world phenomena in machine learning and mechanism design

Accounting for real world phenomena in machine learning and mechanism design
Accounting for real world phenomena in machine learning and mechanism design
As data becomes more readily available, individuals and organisations are increasingly relying on automated systems to make decisions on their behalf. Both machine learning and mechanism design play key roles in the design of such systems. Machine learning is often deployed to learn complex decision rules that mimic or improve upon those adopted by humans. Meanwhile, mechanism design is often deployed to ensure that decision rules satisfy certain axiomatic properties of interest to the designer, such as fairness and incentive compatibility. Unfortunately, many real world settings fall outside the scope of traditional machine learning and mechanism design frameworks. This thesis investigates how approaches from mechanism design and machine learning can be rigorously extended and adapted for such settings to yield meaningful theoretical guarantees.

In particular, we investigate three problem domains; 1) linear regression in the presence of strategic agents, 2) sequential resource deployment with reusable resources and 3) repeated matching with reusable resources. For the first problem domain, we provide a theoretical framework based on Stackelberg predictions games. When the incentives of agents can be captured by a square loss function, we provide a polynomial time algorithm minimising Stackelberg risk, a natural analog to risk in classical supervised learning. For the second problem domain, we introduce a new multi-armed bandit model, called the adversarial blocking bandit problem, which incorporates nonstationary reward sequences and resource unavailability. In particular, we provide finite-time regret guarantees for this setting, by benchmarking against an oracle algorithm which approximates the optimal arm pulling policy. Lastly, for the third problem domain, we introduce a new sequential matching setting, in which a central planner is tasked with constructing matchings repeatedly through time under the assumption that some goods or services may become temporarily unavailable once assigned. Motivated by the random serial dictatorship algorithm, we construct an algorithm for the setting which is approximately truthful and approximately maximises social welfare.
University of Southampton
Bishop, Nicholas
e2b8dc1a-a609-4709-84af-9b2455fd73e6
Bishop, Nicholas
e2b8dc1a-a609-4709-84af-9b2455fd73e6
Gerding, Enrico
d9e92ee5-1a8c-4467-a689-8363e7743362
Tran-Thanh, Long
e0666669-d34b-460e-950d-e8b139fab16c

Bishop, Nicholas (2023) Accounting for real world phenomena in machine learning and mechanism design. University of Southampton, Doctoral Thesis, 179pp.

Record type: Thesis (Doctoral)

Abstract

As data becomes more readily available, individuals and organisations are increasingly relying on automated systems to make decisions on their behalf. Both machine learning and mechanism design play key roles in the design of such systems. Machine learning is often deployed to learn complex decision rules that mimic or improve upon those adopted by humans. Meanwhile, mechanism design is often deployed to ensure that decision rules satisfy certain axiomatic properties of interest to the designer, such as fairness and incentive compatibility. Unfortunately, many real world settings fall outside the scope of traditional machine learning and mechanism design frameworks. This thesis investigates how approaches from mechanism design and machine learning can be rigorously extended and adapted for such settings to yield meaningful theoretical guarantees.

In particular, we investigate three problem domains; 1) linear regression in the presence of strategic agents, 2) sequential resource deployment with reusable resources and 3) repeated matching with reusable resources. For the first problem domain, we provide a theoretical framework based on Stackelberg predictions games. When the incentives of agents can be captured by a square loss function, we provide a polynomial time algorithm minimising Stackelberg risk, a natural analog to risk in classical supervised learning. For the second problem domain, we introduce a new multi-armed bandit model, called the adversarial blocking bandit problem, which incorporates nonstationary reward sequences and resource unavailability. In particular, we provide finite-time regret guarantees for this setting, by benchmarking against an oracle algorithm which approximates the optimal arm pulling policy. Lastly, for the third problem domain, we introduce a new sequential matching setting, in which a central planner is tasked with constructing matchings repeatedly through time under the assumption that some goods or services may become temporarily unavailable once assigned. Motivated by the random serial dictatorship algorithm, we construct an algorithm for the setting which is approximately truthful and approximately maximises social welfare.

Text
Thesis_3au - Version of Record
Available under License University of Southampton Thesis Licence.
Download (1MB)
Text
Final-thesis-submission-Examination-Mr-Nicholas-Bishop
Restricted to Repository staff only

More information

Published date: 4 February 2023

Identifiers

Local EPrints ID: 474341
URI: http://eprints.soton.ac.uk/id/eprint/474341
PURE UUID: 622223bf-5007-4973-adb2-d92789d63809
ORCID for Nicholas Bishop: ORCID iD orcid.org/0000-0001-7062-9072
ORCID for Enrico Gerding: ORCID iD orcid.org/0000-0001-7200-552X
ORCID for Long Tran-Thanh: ORCID iD orcid.org/0000-0003-1617-8316

Catalogue record

Date deposited: 20 Feb 2023 17:53
Last modified: 17 Mar 2024 03:03

Export record

Contributors

Author: Nicholas Bishop ORCID iD
Thesis advisor: Enrico Gerding ORCID iD
Thesis advisor: Long Tran-Thanh ORCID iD

Download statistics

Downloads from ePrints over the past year. Other digital versions may also be available to download e.g. from the publisher's website.

View more statistics

Atom RSS 1.0 RSS 2.0

Contact ePrints Soton: eprints@soton.ac.uk

ePrints Soton supports OAI 2.0 with a base URL of http://eprints.soton.ac.uk/cgi/oai2

This repository has been built using EPrints software, developed at the University of Southampton, but available to everyone to use.

We use cookies to ensure that we give you the best experience on our website. If you continue without changing your settings, we will assume that you are happy to receive cookies on the University of Southampton website.

×