My research interests are in the development of applied decision science and data science
capabilities with a focus on domains such as healthcare, legal and public services,
manufacturing, and management. Most of my work is interdisciplinary and collaboratively
involving both academic colleagues from different fields (e.g. Computer Science, Music,
Sports, Forensics) and external partners including companies from different sectors, the government (IPA),
and public (e.g. NHS Digital and hospitals) and legal services (e.g. Weightmans).
I have generated research income worth a total of more than £1m attracting funding
from sources such as ESRC, Cabinet Office (IPA), EPSRC, Innovate UK, HSIF, and external partners.
Below is a summary of research streams I have been pursuing since 2008, which is when I started with my PhD studies.
Non-standard challenges in experimental optimization
Many of the work packages discussed in this research stream have been initiated during my PhD studies
and are covered in more or less detail in my PhD thesis.
- Ephemeral resource constraints (ERCs): Together with Joshua Knowles I have been
working on a type of dynamic constraints that simulated the availability of resources
required in the evaluation of a solution. Here, evaluations refer to physical experiment
and/or a time-consuming simulation. For example, using a particular instrument setting
may commit us to using this setting for a certain amount of time, or a machine may break down
and thus become unavailable termporarily, or an expert may be required to operate with certain
instrument settings (e.g. due to safety concerns) and she may be available on certain days of
the week only. ERCs define the set of solutions evaluable at any point in time but they
do not change the global optima. We introduced
and formally defined different ERCs based
on our own observations in collaborative experimental work,
analyzed some ERCs from
point of view using Markov Chains, proposed and analyzed different
static strategies and
strategies for dealing with ERCs, and
optimizers with on-line purchasing strategies for resources.
- Objectives with non-uniform latencies: In experimental multiobjective optimization,
it does not need to be the case that all objectives take equal amount of time to be evaluated.
For instance, developing a new washing powder that is cost-efficient and performs well may require
quick evaluations of the cost (summing up the ingredients) but time-consuming washing cycles at
different temperatures, loads, and clothing types. I have kicked off my work on this topic with
Joshua Knowles introducing
non-uniform loosly and proposing and analyzing simply strategies. Julia Handl
then joined us in a follow up paper where we provided a
introduction to non-uniform
latencies, and proposed and analyzed a set of new strategies. Together with Kaisa Miettinen, Tinkle Chugh,
and Vesa Ojalehto, I have now also proposed and analyzed
methods to deal with non-uniform latencies in objectives.
- Changing decision variables during search: The significant demand in resources (e.g. time,
money, labour) combined with time pressure to deliver results can lead to situations where an
experimentalist wants to change the objects being optimized, and since restarting optimization would
be too expensive and wasteful (in many cases) it is desired to do this on the fly (i.e. during the
optimization). For instance, in a drug discovery problem where one aims to discover potent combinations
of drugs drawn from a drug library, one way want to add, replace of remove drugs from the drug library.
I have been looking at precisely this scenario together with Joshua Knowles and colleagues from the
Manchester Institute of Biotechnology (MIB) including the
current President and Vice-Chancellor of the University of Manchester, Dame Nany Rothwell, and the former
CEO of BBSRC, Douglas Kell. Our work has resulted in an experimental/methodoligical-focussed paper, published
in Nature Chemical Biology,
and an optimization-focussed paper.
- Optimization in lethal environments: In this work we looked at scenarios where the evaluation of a
lethal solution (modelled here as one below a certain fitness threshold) causes the solution to be immediately
removed from the population and the population size to be reduced by one. This models certain closed-loop
evolution scenarios that may be encountered, for example, when evolving
nano-technologies or autonomous robots that are limited in numbers, thus destroying a robot means to lose it for good.
Together with Joshua Knowles we have motivated such scenarios, and proposed and analyzed methods to
cope in lethal environments.
Multiobjective optimization and MCDM
- Accounting for user preferences in multiobjective optimization: Together with
Juergen Branke and Xiaodong Li, I have been working on extending multiobjective particle
swarm optimization (PSO) to account for user preferences during search. We looked at articulating
user preferences in terms of reference points in the objective space, and used a biased steady-state approach within PSO
to ensure that the optimizer discovers areas in the objective space closest to the reference points
only. This work formed the body of my Diplom dissertation, and our
paper was one of the first that
combined user preferences with multiobjective PSO.
- Multiobjective optimization with non-uniform latencies in objectives: Together with Josh
I looked at problems where the evaluation times vary across objectives in a multiobjective problem.
We looked at this issue in the context of experimental optimization. More information and links to my
publications on this topic are provided above in the research stream on
in experimental optimization.
- Multiobjective optimization of instrument operating conditions: Together with Nigel J. Titchener-Hooker,
Suzy Farid, and Spyridon Gerontas, I have investigated
the application of multiobjective optimization to
the optimization of chromatographic operating conditions in the context of biopharmaceutical production.
Chromatography is an established but costly laboratory technique to separate (purify) a mixture,
such as drug mixture, from impurities.
- Navigation in multiobjective optimization: Together
Matthias Ehrgott, Xavier Gandibleux, Martin Josef Geiger,
Kathrin Klamroth and Mariano Luque, I have formally defined and reviewed an approach towards a common understanding of
search and decision-making strategies to identify the most preferred solution among the Pareto set for a multiobjective
optimization problem. We refer to this approach as navigation,
and methods adopting this approach faciliate the decision maker
to interactively learn about the problem. In contrast, a decision support system learns about the preferences of the decision
maker. This work is based on the work we did as a group at the
Dagstuhl Seminar 12041, Learning in Multiobjective Optimization, 2012.
- Review of surrogate-assisted multicriteria optimization: Together with Michael Emmerich, Jussi Hakanen,
Yaochu Jin, and Enrico Rigoni, I have reviewed
emerging complexity-related topics and challenges in surrogate-assisted
multicriteria optimization that may not be prevalent in nonsurrogate-assisted single-objective optimization.
We discuss several promising future research directions and prospective solutions to tackle these challenges, and,
finally, we provide insights from an industrial point of view
into how surrogate-assisted multicriteria optimization techniques can be developed and applied within a collaborative
business environment to tackle real-world problems.This work is based on the work we did as a group at the
Dagstuhl Seminar 15031, Understanding
Complexity in Multiobjective Optimization, 2015.
- Constraint-handling in surrogate-assisted optimization: Together with Samineh Bagheri, Wolfgang Konen, Jeurgen Branke,
Kalyanmoy Deb, Jonathan Fieldsend, Domenico Quagliarella, and Karthik Sindhya,
I have developed an approach to deal with constraints
in surrogate-assisted optimization. In particular, we introduced a new EGO-based algorithm which tries to overcome some of the common issues
with Kriging optimization algorithms when applied to constrained problems: (1) early stagnation, (2) problems with multiple active
constraints and (3) frequent crashes. We have validated our algorithm on both synthetic problems (G-function suite) and on an airfoil shape example.
This work is based on the work we did as a group at the
Lorentz Centre Seminar SAMCO: Surrogate-Assisted Multi-Criteria
- Adaptive encoding for evolutionary data clustering: Together with my PhD student Cameron Shand, Julia Handl
and John Keane, I have been looking at adaptive encodings for optimization approaches applied to data clustering. The
motivation behind this research is to create optimization algorithms that are capable of adjusting autonomously to
dataset at hand. In particular, our research investigated the adaption of a specific hyperparameter - which directly governs
the encoding granularity - in the multi-objective clustering algorithm MOCK. Our proposed approach leads to a ~40% reduction
of computational expense while achieving clusterint performance that is at least as good as achieved for an optimal (a priori) setting
of the hyperparameter. Our paper can be downloaded here.
- Clustering problem generator: Together with my PhD student Cameron Shand, Julia Handl
and John Keane, I am currently working on evolving controllably difficult datasets for clustering.
Drug development and manufacture
During my Postdoc at the Biochemical Engineering Department at UCL (2011 - 2015), I had the opportunity to work on a range of
interesting problems related to the development and manufacture of drugs. I continue to work on these and other related topics
if the opportunity arises. Below is an overview of my work in this research stream.
- Optimization of drug manufacturing processes: Increasing demand in therapeutic drugs has
resulted in the need to design cost-effective, flexible and robust manufacturing processes capable of
meeting regulatory product purity requirements. Together with Suzy Farid and Sofia Simaria, I have
designed a framework linking an
with a biomanufacturing process economics model to discover manufacturing processes that provide the
best trade-off with respect to cost of goods per gram (COG/g), robustness in COG/g, and impurity
removal capabilities. The framework also simulates and optimizes subject to various process
uncertainties and design constraints. Our work has been pubished in
the computer science community, where
the focus was on the methodology and the optimization techniques used,
and the bioprocessing community,
where the focus was rather on demonstrating the usefulness of our a decision support
tool from an application perspective. Our earlier work with Richard Turner (MedImmune) looked at
the single-objective case (Cost
of Goods was the objective) of optimizing drug manufacturing processes.
- Lot sizing and scheduling of multi-site drug production: Together with my PhD student Folarin Oyebolu, Jeroen van Lidth de Jeude,
Cyrus Siganporia, Suzy Farid, and Juergen Branke, I have been looking at
production planning for large biopharmaceutical companies.
In such cases, planning typically needs to done across multiple products and several production facilities.
Production is also usually done in batches with a substantial set-up cost and time for switching between products. The objective in
this planning problem is to satisfy demand while minimising manufacturing, set-up and inventory costs. The resulting production planning problem
is a compabinatorial optimization problem that is a variant of the capacitated lot-sizing and scheduling problem,
We proposed a tailored construction heuristic that schedules demands
of multiple products sequentially across several facilities to build a multi-year production plan (solution). The sequence in which
the construction heuristic schedules the different demands was optimised by a genetic algorithm. Our approach is able to
outperform a mathematical programming model for certain scenarios because the discretisation of
time in mathematical programming artificially restricts the solution space.
- Life-cycle and cost of goods assessment of different drug manufacturing technologies: Together with my MSc student Phumthep Bunnak,
Sri V. Ramasamy, Paola Lettieri, and Nigel J. Titchener-Hooker, I have looked at the application of life-cycle assessment (LCA) to
the bioprocess sector in order to contribute toward the design of more cost-efficient, robust and environmentally-friendly manufacturing processes.
In particular, we compared the two most commonly used upstream configurations for mAb manufacture, namely fed-batch
(FB) and perfusion-based processes. We have shown that
LCA in combination with a cost analysis can be a powerful tool
when it comes to technology selection in the bioprocess sector.
In addition to drug development and manufacture, I am working on a range of other applications: