49. OR in the Age of AI

In 2017, I was awarded Honorary Membership of the Operational Research Society. I felt duly honoured, not least because I had considered myself, in part, an operational researcher since the 1970s and had indeed published in the Society’s Journal and was a Fellow at a time when that had a different status. However, there was a price! For the following year, I was invited to give the annual Blackett Lecture, delivered to a large audience at the Royal Society in November last year. The choice of topic was mine. Developments in data science and AI are impacting most disciplines, not least OR. I thought that was something to explore and that gave me a snappy title: OR in the Age of AI.

OR shares the same enabling disciplines as data science and AI and (in outline) is concerned with system modelling, optimisation, decision support, and planning and delivery. The systems focus forces interdisciplinarity and indeed this list shows that insofar as it is a discipline, it shares its field with many others. If we take decision support and planning and delivery as at least in part distinguishing OR, then we can see it is applied and it supports a wide range of customers and clients. These have been through three industrial revolutions and AI promises a fourth. We can think of these customers, public or private, as being organisations driven by business processes. What AI can do is read and write, hear and see, and translate, and these wonders will transform many of these business processes. There will be more complicated shifts – driving robotics, including soft robotics, understanding markets better, using rules-based algorithms to automate processes – some of them large-scale and complicated. In many ways all this is classic OR with new technologies. It is ground-breaking, it is cost saving, and does deplete jobs. But in some ways it’s not dramatic.

The bigger opportunities come from the scale of available data, computing power and then two things: the ability of systems to learn; and the application to big systems, mainly in the public sector, not driven in the way that profit-maximising industries are. For OR, this means that the traditional roles of its practitioners will continue, albeit employing new technologies; and there is a danger that because these territories overlap across many fields – whether in universities or the big consultancies – there will be many competitors that could shrink the role of OR. The question then is:  can OR take on leadership roles in the areas of the bigger challenges?

Almost every department of Government has these challenges – and indeed many of them – say those associated with the criminal justice system – embrace a range of government departments, each operating in their own silos, not combining to collect the advantages that could be achieved if they linked their data. They all have system modelling and/or ‘learning machine’ challenges. Can OR break into these?

The way to break in is through ambitious proof-of-concept research projects – the ‘R’ part of R and D – which then become the basis for large scale development projects, the ‘D’. There is almost certainly a systemic problem here. There have been large scale ambitious projects – usually concerned with building data systems – arguably a prerequisite – and many of these fail. But most of the funded research projects are relatively small and the big ‘linking’ projects are not tackled. So the challenge for OR, for me, is to open up to the large-scale challenges, particularly in government, to ‘think big’.

The OR community can’t do this alone, of course. However, there is a very substantial OR service in government – one of the recognised analytics professions – and there is the possibility of asserting more influence from within. But the Government itself has a responsibility to ensure that its investment in research is geared to meet these challenges. This has to be a UKRI responsibility – ensuring that research council money is not too thinly spread, and ensuringthat research councils work effectively together as most of the big challenges are interdisciplinary and cross-council. Government Departments themselves should both articulate their own research challenges and be prepared to fund them.

Alan Wilson

48. Mix and match: The five pillars of data science and AI

There are five pillars of data science and AI. Three make up, in combination, the foundational disciplines – mathematics, statistics and computer science; the fourth is the data – ‘big data’ as it now is; and the fifth is a many-stranded pillar – domain knowledge. The mathematicians use data to calibrate and test models and theories; the statisticians also calibrate models and seek to infer findings from data; the computer scientists develop the intelligent infrastructure (cf Blog 47). Above all, the three combine in the development of machine learning – the heart of contemporary AI and its applications. Is this already a new discipline? Not yet, I suspect – not marked by undergraduate degrees in AI (unlike, say, biochemistry). These three disciplines can be thought of as enabling disciplines and this helps us to unpick the strands of the fifth pillar: both scientists and engineers are users, as are the applied domains such as  medicine, economics and finance, law, transport and so on. As the field develops, the AI and data science knowledge will be internalised in many of these areas – in part meeting the Mike Lynch challenge (see Blog 46) incorporating prior knowledge into machine learning.

Even this brief introduction demonstrates that we are in a relatively new interdisciplinary field. It is interesting to continue the exploration by connecting to previous drivers of interdisciplinarity – to see how these persist and ‘add’ to our agenda; and then to examine examples of new interdisciplinary challenges.

It has been argued in earlier posts that the concept of a system of interest drives interdisciplinarity and this is very much the case here in the domains for which the AI toolkit is now valuable. More recently, complexity science was an important driver with challenges articulated through Weaver’s notion of ‘systems of organised complexity’. This emphasises both the high dimensionality of systems of interest and the nonlinear dynamics which drives their evolution. There are challenges here for the applications of AI in various domains. Handling ‘big data’ also drives us towards high dimensionality. I once estimated the number of variables I would like to have to describe a city of a million people at a relatively coarse grain, and the answer came out as 1013! This raises new challenges for the topologists within mathematics: how to identify structures within the corresponding data sets – a very sophisticated form of clustering! These kinds of system can be described through conditional probability distributions again with large numbers of variables – high dimensional challenges for Bayesian statisticians. One way to proceed with mathematical models that are high dimensional and hence intractable is to run them as simulations. The outputs of these models can then be treated as ‘data’ and, to my knowledge, there is an as-yet untouched research challenge: to apply unsupervised machine learning algorithms to these outputs to identify structures in a high-dimensional nonlinear space.

We begin to reveal many research challenges across both foundational, and especially, applied domains. (In fact a conjecture is that the most interesting foundational challenges emerge from these domains?) We can then make another connection – to Brian’s Arthur’s argument in his book The nature of Technology. A discovery in one domain can, sometimes following a long period, be transferred into other domains: opportunities we should look out for.

Can we optimise how we do research in data science and AI? We have starting points in the ideas of systems analysis and complexity science: define a system of interest and recognise the challenges of complexity. Seek the data to contribute to scientific and applied challenges – not the other way round – and that will lead to new opportunities? But perhaps above all, seek to build teams which combine the skills of mathematics, statistics and computer science, integrated through both systems and methods foci. This is non-trivial, not least due to the shortage of these skills. In the projects in the Turing Institute funded by the UKRI Special Priorities Fund – AI for Science and Government (ASG) and Living with machines (LWM) – we are trying to do just this. Early days and yet to be tested. Watch this space!

Alan Wilson

47. What is ‘data science? What is ‘AI’?

When I first took on the role of CEO at The Alan Turing Institute, the strap line beneath the title was ‘The National Institute for Data Science’. A year or so later, this became ‘The National Institute for Data Science and AI’ – at a time when there was a mini debate about whether there should be a separate ‘national institute for AI’. It has always seemed to me that ‘AI’ was included in ‘data science’ – or maybe vice versa. In the early ‘data science’ days, there were plenty of researchers in Turing focused on machine learning for example. However, we acquired the new title – ‘for avoidance of doubt’ one might say – and it now seems worthwhile to unpick the meanings of these terms. However we define them, there will be overlaps but by making the attempt, we can gain some new insights.

Ai has a long history, with well-known ‘summers’ and ‘winters’. Data science is newer and is created from the increases in data that have become available (partly generated  by the Internet of Things) closely linked with continuing increases in computing power. For example, in my own field of urban modelling, where we need location data and flow data for model calibration, the advent of mobile phones means that there is now a data source that locates most of us at any time – even when phones are switched off. In principle, this means that we could have data that would facilitate real-time model calibration. New data, ‘big data’, is certainly transforming virtually all disciplines, industry and public services.

Not surprisingly, most universities now have data science (or data analytics) centres or institutes – real or virtual. It has certainly been the fashion but may now be overtaken by ‘AI’ in that respect. In Turing, our ‘Data science for science’ theme has now transmogrified into ‘AI for science’ as more all embracing. So there may now be some more renaming!

Let’s start the unpicking. ‘Big data’ has certainly invigorated statistics. And indeed, the importance of machine learning within data science is a crucial dimension – particularly as a clustering algorithm with obvious implications for targeted marketing (and electioneering!). Machine learning is sometimes called ‘statistics reinvented’! The best guide to AI and its relationship to data science that I have found is Michael Jordan’s blog piece ‘Artificial intelligence – the revolution hasn’t happened yet’ – googling the title takes you straight there. He notes that historically AI stems from what he calls ‘ human-imitative’ AI; whereas now, it mostly refers to the applications of machine learning – ‘engineering’ rather than mimicking human thinking. As this has had huge successes in the business world and beyond, ‘it has come to be called data science’ – closer to my own interpretation of data science, but which, as noted, fashion now relabels as AI.

We are a long way from machines that think and reason like humans. But what we have is very powerful. Much of this augments human intelligence, and thus, following Jordan, we can reverse the acronym: ‘IA’ is ‘intelligence augmentation’ – which is exactly where the Turing Institute works on rapid and precise machine-learning-led medical diagnosis – the researchers working hand in hand with clinicians. Jordan also adds another acronym: ‘II’ – ‘intelligent infrastructure’. ‘Such infrastructure is  beginning to make its appearance in domains such as transportation, medicine, commerce and finance, with vast implications for individual humans and societies.’ This is a bigger scale concept than my notion that an under-developed field of research is the design of (real-time) information systems.

This framework, for me, provides a good articulation of what AI means now – IA and II. However, fashion and common useage will demand that we stick to AI! And it will be a matter of personal choice whether we continue to distinguish data science within this!!

Alan Wilson

46. Moving on 2: pure vs applied

In my early days as CEO in Turing, I was confronted with an old challenge: pure vs applied though often in a new language – foundational vs consultancy for example. In my own experiences from my schooldays onwards, I was always aware of the higher esteem associated with the ‘pure’ and indeed leaned towards that end of the spectrum. Even when I started working in physics, I worked in ‘theoretical physics’. It was when I converted to the social sciences that I realised that my new fields, I could have it both ways: I worked on the basic science of cities through mathematical and computer modelling but with outputs that were almost immediately applicable in town and regional planning. So where did that kind of thinking leave me in trying to think through a strategy for the Institute?

Oversimplifying: there were two camps – the ‘foundational’ and the ‘domain-based’. Some of the former could characterise the latter as ‘mere consultancy’. There were strong feelings. However, there was a core that straddled the camps: brilliant theorists, applying their knowledge in a variety of domains. It was still possible to have it both ways. How to turn this into a strategy – especially given that the root of a strategic plan will be the allocation of resources to different kinds of research? In relatively early days, it must have been June 2017, we had the first meeting of our Science Advisory Board and for the second day, we organised a conference, inviting the members of our Board to give papers. Mike Lynch gave a brilliant lecture on the history of AI through its winters and summers with the implicit question: will the present summer be a lasting one? At the end of his talk, he said something which has stuck in my mind ever since: “The biggest challenge for machine learning is the incorporation of prior knowledge”. I would take this further and expand ‘knowledge’ to ‘domain knowledge’. My intuition was that the most important AI and data science research challenges lay within domains – indeed that the applied problems generated the most challenging foundational problems.

Producing the Institute’s Strategic Plan in the context of a sometimes heated debate was a long drawn out business – taking over a year as I recall. In the end, we had a research strategy based on eight challenges, six of which were located in domains: health, defence and security, finance and the economy, data-centric engineering, public policy and what became ‘AI for science’. We had two cross-cutting themes: algorithms and computer science, and ethics. The choice of challenge areas was strongly influenced by our early funders:  the Lloyds Register Foundation, GCHQ and MoD, Intel and HSBC. Even without a sponsor at that stage, we couldn’t leave out ‘health’! All of these were underpinned by the data science and machine learning methods tool kit. Essentially, this was a matrix structure: columns as domains, rows as method – an effective way of relaxing the tensions, of having it both ways. This structure has more or less survived, though with new challenges added – ‘cities’ for example and the ‘environment’.

When it comes to allocating resources, other forces come into play. Do we need some quick wins? The balance between the short term and the longer – the latter inevitably more speculative? Should industry fund most of the applied? This all has to be worked in the context of a rapidly developing Government research strategy (with the advent of UKRI) and the development of partnerships with both industry and the public sector. There is a golden rule, however, for a research institute (and for many other organisations such as universities): think through your own strategy rather than simply ‘following the money’ which is almost always focused on the short term. Then given the strategy, operate tactically to find the resources to support it.

In making funding decisions, there is an underlying and impossible question to answer: how much has to be invested in an area to produce results that are truly transformative? This is very much a national question but there is a version of it at the local level. Here is a conjecture: that transformative outcomes in translational areas demand a much larger number of researchers to be funded than to produce such transformations in foundational areas. This is very much for the ‘research’ end of the R and D spectrum – I can see that the ‘D’ – development – can be even more expensive. So what did we end up with? The matrix works and at the same time acknowledges the variety of viewpoints. And we are continually making judgements about priorities and the corresponding financial allocations. Pragmatism kicks in here!

Alan Wilson

37: The ‘Leicester City’ phenomenon: aspirations in academia.

Followers of English football will be aware that the top tier is the Premier League and that the clubs that finish in the top four at the end of the season play in the European Champions’ League in the following year. These top four places are normally filled by four of a top half a dozen or so clubs – let’s say Manchester United, Manchester City, Arsenal, Chelsea, Tottenham Hotspur and Liverpool. There are one or two others on the fringe. This group does not include Leicester City. At Christmas 2014, Leicester were bottom of the Premier League with relegation looking inevitable. They won seven of their last nine games in that season and survived. At the beginning of the current (2015-16) season, the bookmakers’ odds on them winning the Premier League were 5000-1 against. At the time of writing, they top the league by eight points with four matches to play. The small number of people who might have bet £10 or more on them last August are now sitting on a potential fortune.

How has this been achieved? They have a very strong defence and so concede little; they can score ‘on the break’, notably through Jamie Vardy, a centre forward who not long ago was playing for Fleetwood Town in the nether reaches of English football; they have a thoughtful, experienced and cultured manager in Claudio Ranieri; and they work as a team. It is certainly a phenomenon and the bulk of the football-following population would now like to see them win the League.

What are the academic equivalents? There are university league tables and it is not difficult to identify a top half dozen. There are tables for departments and subjects. There is a ranking of journals. I don’t think there is an official league table of research groups but certainly some informal ones. As in football, it is very difficult to break into the top group from a long way below. Money follows success – as in the REF (the Research Excellence Framework) – and facilitates the transfer of the top players to the top group. So what is the ‘Leicester City’ strategy for an aspiring university, an ambitious department or research group, or a journal editor? The strong defence must be about having the basics in place – good REF ratings and so on. The goal-scoring break-out attacks is about ambition and risk taking. The ‘manager’ can inspire and aspire. And the team work: we are almost certainly not as good as we should be in academia, so food for thought there.

Then maybe all of the above requires at the core – and I’m sure Leicester City have these qualities – hard work, confidence, and good plans while still being creative; and a preparedness to be different – not to follow the fashion. So when The Times Higher has its ever-expanding annual awards, maybe they should add a ‘Leicester City Award’ for the university that matches their achievement in our own leagues. Meanwhile, will Leicester win the League? Almost all football followers in the country are now on their side. We will see in a month’s time!

Alan Wilson, April 2016

27: Beware of optimisation

The idea of ‘optimisation’ is basic to lots of things we do and to how we think. When driving from A to B, what is the optimum route? When we learn calculus for the first time, we quickly come to grips with the maximisation and minimisation of functions. This is professionalised within operational research. If you own a transport business, you have to plan a daily schedule of collections and deliveries. Continue reading “27: Beware of optimisation”

22: Requisite Knowledge

W Ross Ashby was a psychiatrist who, through books such as Design for a brain, was one of the pioneers of the development of systems theory (qv) in the 1950s. A particular branch of systems theory was ‘cybernetics’ – from the Greek ‘steering’ – essentially the theory of the ‘control’ of systems. This was, and I assume is, very much a part of systems engineering and it attracted mathematicians such as Norbert Weiner. For me, an enduring contribution was ‘Ashby’s Law of Requisite Variety’ which is simple in concept and anticipates much of what we now call complexity science. ‘Variety’ is a measure of the complexity of a system and is formally defined as the number of possible ‘states’ of a system of interest. A coin to be tossed has two possible states – heads or tails; a machine can have billions. Continue reading “22: Requisite Knowledge”