Sign Out
Logged In:
Tab Image

OR58 - Plenary Speakers

Graham Fletcher

The Changing Face of the Transport Problem.

Transport is a changing problem. Governments and cities used to invest in infrastructure, a road or a railway, an airport or a station. We used to use well understood mathematical models of how, why and when people travel to predict usage and understand the value and impact of these investments.

Unfortunately, the way people think about travel, buy travel and their travel options are changing. Perhaps the biggest change that the industry sees is that the "systems integrator" is changing. The travelling public used to take all the components of the travel system and integrate them to provide a solution. But that expectation is changing, many organisations, both private and public, are now offing to integrate the system for you, to work out how you should travel, when and to arrange the tickets.

The problem with integration being moved from the traveller to a relatively small number of people is that those people have a huge power over the system. Consider what would happen if everybody's smart phone recommended that they take the same road as opposed to spreading the load, or how many autonomous vehicles you would need to control to effectively change the speed limit in a city, irrespective of what the traffic planners want to happen.

In this talk Graham will outline what he sees as the future challenges to our community raised by these changing transportation problems.

Short biography

Graham is head of research and development for Cubic Transportation Systems ltd. CTS are a worldwide leader in delivering integrated transport solutions in major cities. Indeed while most of you won’t know it, most of you will have used a CTS product. Late last year CTS committed £m's to increasing its research footprint, including opening a new innovation centre in London. Graham now leads a worldwide team that collaborates with academics around the globe from the new facility. One of Graham's close collaborations is with the University of Portsmouth, our hosts.

Until very recently Graham was the Modelling and Visualisation Director at Transport Systems Catapult, one of a network of elite technology and innovation centres established by Innovate UK as a long-term investment in the UK’s economic capability. He joined Transport Systems Catapult in January of 2013 from the defence sector.  However, he advises that much of his future will still be in the transport sector.  Graham graduated from Royal Holloway, University of London in 1991 with a B.Sc. in Computer Science and continued his education at Loughborough University qualifying as a PhD in Non-linear Optimisation and Knowledge Elicitation.

Ruth Kaufman

Diversity and Identity: Challenges and Opportunities for UK O.R.
People engaged in an activity like O.R., can be described on countless dimensions. For example, personal characteristics (the 'protected characteristics' of ethnicity, gender etc that are part of formal diversity initiatives, but also values, preferences, personality), daily work activities, customer, owner, career path or whether they have ever heard of O.R.  Diversity can be a strength, but so can uniform identity. This talk reviews some significant dimensions of diversity and identity in UK O.R. and considers what we might need to do to overcome the challenges of too much or too little diversity and where we can exploit the enormous potential benefits of the glorious variety of ways of 'being an O.R. person'.  (A more detailed abstract will follow nearer the event).

Short biography

Ruth Kaufman, a ‘Companion’ of the OR Society, became President of the Society in January of this year.  Like many other members of the OR Society, she fell into O.R. by accident, in her case having taken a maths BA in the School of Social Sciences at Sussex University.  This led to a long career in public sector O.R. and wider management at London Transport, London Electricity, Department of Health and Export Credits Guarantee Department (ECGD).  At ECGD, she joined the Executive Board, having responsibility for strategy and change management as well as leading an influential OR group. In the voluntary sector, Ruth chaired a small charity, Woman's Trust, for five years and was a founder member of the OR Society’s Pro Bono Scheme. Ruth is currently (amongst other things), Advisor to the Finance Committee at the National Federation of Women’s Institutes, a visiting Senior Fellow at the London School of Economics and a freelance consultant and advisor. She was awarded an OBE for 'services to Operational Research' in the 2016 New Year's Honours list.

During her two-year presidency of the OR Society, Ruth aims to help O.R. become more visible and to help the OR Society grow its reach and impact, nurture OR research and build the people pipeline.

Peter Richtárik

Introduction to Big Data Optimization
Randomized coordinate descent (RCD) methods are algorithms of choice for many practical optimization problems arising in engineering, computer science, statistics, applied mathematics and machine learning. They are especially well suited for solving big data problems – problems described by huge quantities of data, often involving millions or billions of variables (“coordinates"). A basic variant of RCD in each iteration updates a single coordinate of the decision vector, chosen uniformly at random. This is done either by performing a one-dimensional minimization or, preferably, by applying a simple closed-form formula. RCD methods can therefore be seen as iterative randomized decomposition techniques, which operate by reducing a large-dimensional problem into a sequence of randomly generated one-dimensional problems. A successful and popular strategy for speeding up RCD, both in theory and in practice, is to sample the coordinates with specific problem-dependent non-uniform probabilities. Another very successful strategy is to update a random subset of coordinates of a fixed size instead, possibly in parallel or in a distributed environment. The latter strategy is very closely related to a popular machine learning technique known as mini-batching. RCD methods can be accelerated in the sense of Nesterov, which leads to further theoretical and practical benefits. 

In this talk I will give an accessible introduction into the field of randomized coordinate descent algorithms from a modern perspective for which we coined the phrase “arbitrary sampling”. The methods I shall describe can at every iteration pick and update a random subset of the coordinates (by design, this can be done in parallel), chosen in an i.i.d. fashion according to an arbitrary random set-valued mapping (aka arbitrary sampling). That is, one may in theory assign a unique probability to each of the exponentially many subsets of the set of coordinates, and pick and update coordinates in this manner. With this approach, the line separating the world of randomized and deterministic methods is fully removed – and what emerges is a unified theory including both randomized and deterministic first-order algorithms in special cases. I shall describe standard, accelerated and primal-dual variants developed in my group – with the latter being a state-of-the-art method for training linear predictors. 

Short biography

Peter Richtárik is an Associate Professor in the School of Mathematics, University of Edinburgh. He obtained his PhD from Cornell University in 2007, after which he spent two years as a postdoctoral fellow at Universite Catholique de Louvain. Recently, he has been a visiting researcher at the Simons Institute at UC Berkeley. He is a Faculty Fellow at the Alan Turing Institute – the UK national research centre for data science. Prof Richtárik has made fundamental contributions to “big data optimization”, which is an emerging field at the intersection of mathematical optimization, convex analysis, probability theory, computer science, machine learning and high performance computing. He is the recipient of an EPSRC Fellowship in the Mathematical Sciences, which allows him to focus on further developing this line of work. In a sequence of influential papers, mostly written in collaboration with his PhD students and postdocs, he developed the theory of randomized coordinate descent and stochastic gradient methods of various flavours for convex optimization in very high dimensions. These algorithms have substantially better complexity bounds than previous deterministic methods, and have state-of-the-art practical performance for key statistical learning tasks involving large data. The underlying complexity theory is built on a blend of ideas and tools from optimization, convex analysis, linear algebra, computer science and probability theory. Prof Richtárik is the recipient of the 2016 SIAM SIGEST Paper Award and a 2016 EUSA Best Research or Dissertation Supervisor Award. His PhD students and postdocs have received a number of national and international prizes and awards for joint work, including 17th and 16th IMA Leslie Fox Prize (2015 and 2013), BASP Frontiers Best Contribution Award (2015), Optimization and Big Data Best Contribution Award (2015), OR Society Best Doctoral Dissertation Award (2014), and the INFORMS Computing Society Best Student Paper Prize (2012).