Some recent work at UWS was inspired by the real or imagined activities of a colony of foraging bees. Things got awkward when the model that was being used for experiments turned out to be different from the way bees actually forage. The supervisors said “all models are wrong, but some are useful” and of course wanted to explore the proposed approach, but the student largely lost interest. To my surprise, some readers felt that the study of unnatural systems was intrinsically repugnant, and that the story illustrated the need for science and religion to work hand in hand.
Suppose that a multi-agent system, with the task of looking for a particular sort of cluster in a large data set, observes a potential sub-cluster. We can imagine an automated step of spawning a new agent trained to look for further evidence of such a cluster. However, it is a bit fanciful to think of this new agent as a specially trained infant bee, as bees may learn the habits of the nest, but do not seem to receive the sort of individual instruction found in species with nuclear families. Other work at UWS examined the development of language in interacting groups of automata, and the introduction of a new word in that experiment is not unlike the introduction of a new agent in this one, since the introduction of the word implies a new subset of individuals that use it.
Leaving aside the biological inspiration, could a commercial system be imagined with similar properties? We could imagine such a system working in the data centre of a large supermarket or bank. If new agents can be spawned in this way, there would undoubtedly be issues of monitoring or control. A novel data cluster might result in a massive generation of new agents which might appear as unexpected additional activity in the system. In a commercial data centre it is possible that such an event would lead to suspicions of an intrusion or system fault.
If the autonomous agents are required to do a lot of status reporting to explain what they are up to, the additional monitoring traffic might create so many external messages as to call into question the wisdom of using agents. On the other hand, if the reporting traffic was cleverly aggregated within the swarm, a coherent report could be made to a monitor that a particular observation led to the deployment of 123,000 agents to investigate the possible existence of a new cluster, and this activity had now ended. Some computing systems build this sort of observable surface over chaotic, Brownian, internal motion; just as the apparently random behaviour of autonomous bees creates a regular-shaped nest. For example, network management systems aggregate event reports that have a shared cause, and (doubtless) Microsoft’s performance reporting systems do something similar.
In this way, the investigation of circumstances for and rules for the creation of a new agent leads to a new and interesting control problem, where the new problem is that of explaining the new situation that has arisen, in terms that make sense to those who have not been tracking all the details…
Wednesday, September 9, 2009
Monday, September 7, 2009
On Scholarship 2.0
During August Reinventing Academic Publishing Online appeared on Scholarship 2.0. It is a polemic against what its authors see as an exclusive establishment consisting of the "top academic journals" that only the richest universities can afford, and a self-serving institutional system that distorts the academic process in order to make the job of funding bodies and appointing committees easier.
Now there are many misguided people who think that there are such things as "top academic journals" where the best computing research is to be found, and regrettably some of these people do appear to hold positions of power and influence. But the fault is theirs alone.
I believe strongly in the value of computing research conferences: but the large ones have pursued profit at the expense of discrimination. It is easy to find dreadful papers presented at even the best conferences, with half-baked ideas and without results or any pretence at evaluation. But it is precisely the Web and Web 2.0 that allows us to find quality independently of the vehicle used for publication.
I have been following with some interest the response in UK academia to the recent Research Assessment Exercise. The Computing panel noted that excellent articles were to be found in journals with low impact factors, and conversely. They were astonished at the huge number (1247) of refereed journals that submitted articles had appeared in, and were amazed that relatively few university departments had submitted conference publications. They restated their policy that conference papers could be just as good as those found in the "top-rated journals".
Interested readers can follow this debate in the Conference of Heads and Professors of Computing and the consultation about the Research Excellence Framework.
Not only is their analysis of the last RAE excellent: so are the proposals for the next RAE, which will recognise that originality, rigour and impact are not usually found in a single publication. So let's just get on with the research, and leave the task of re-inventing academic publication to those who have time for it.
(Update 9 Nov: and in the meantime, join www.mendeley.com, which looks like a good Web 2.0 scholarship repository!)
Now there are many misguided people who think that there are such things as "top academic journals" where the best computing research is to be found, and regrettably some of these people do appear to hold positions of power and influence. But the fault is theirs alone.
I believe strongly in the value of computing research conferences: but the large ones have pursued profit at the expense of discrimination. It is easy to find dreadful papers presented at even the best conferences, with half-baked ideas and without results or any pretence at evaluation. But it is precisely the Web and Web 2.0 that allows us to find quality independently of the vehicle used for publication.
I have been following with some interest the response in UK academia to the recent Research Assessment Exercise. The Computing panel noted that excellent articles were to be found in journals with low impact factors, and conversely. They were astonished at the huge number (1247) of refereed journals that submitted articles had appeared in, and were amazed that relatively few university departments had submitted conference publications. They restated their policy that conference papers could be just as good as those found in the "top-rated journals".
Interested readers can follow this debate in the Conference of Heads and Professors of Computing and the consultation about the Research Excellence Framework.
Not only is their analysis of the last RAE excellent: so are the proposals for the next RAE, which will recognise that originality, rigour and impact are not usually found in a single publication. So let's just get on with the research, and leave the task of re-inventing academic publication to those who have time for it.
(Update 9 Nov: and in the meantime, join www.mendeley.com, which looks like a good Web 2.0 scholarship repository!)
Saturday, July 4, 2009
On Truth and Information
In recent years, several philosophers of computing, such as Fred Dretske(1932-) and Luciano Floridi (1964-), have established to a great many people’s satisfaction that for something to be called information, it should allow us to learn something that is true. They argue that false information is not information, just as a false policeman is not a policeman.
In my June 20 contribution to this blog I rather incautiously mentioned the pursuit of the truth about (the laws of) the physical world, and I hinted that I felt this was not the business of science, or research for that matter. Truth is a great concept and a noble ideal, but as I mentioned in the April 18 contribution, the more truthful we make any statement, the less clear it is, and the narrower the scope of its application. In a way, the only absolutely truthful statements are the formal tautologies of pure mathematics, the necessary truths that depend on nothing, and so add nothing to our knowledge. Conversely, any statement that is not necessarily true is merely (by definition) contingently true. It might be true, that is, if (um..) everything were really as it says. There are few philosophers, and even fewer scientists, who feel that this sort of circular discussion is worth while.
And yet, precisely this kind of argument has fascinated people for millennia. Descartes’ necessary truth “cogito, ergo sum” (1641) was the start of his argument to prove the existence of God and the immortality of the soul. A very similar exercise by Bernard Lonergan (1904-1984) achieved a wide currency in 1956. Now the reader should always smell a rat if someone claims to prove a contingent truth from a necessary one. The Australian philosopher David Stove (1927-1994) catalogued a large number of such arguments from Plato to Habermas: explaining that since they want to make everyone accept their opinion, it is a good trick to make it appear merely a logical deduction from a necessary truth. The trick can be made to work, as he explains, with the support of some impressive but contradictory concept, in the same way that division by zero can be used to create a convincing-looking proof. Usually though, such philosophers are vainly trying to use logic to establish some belief which predates their attempt and will outlast their failure.
To return to computing: every piece of information, according to all of us who follow Floridi and Dretske, contains its very own claim of contingent truth. By labelling it as information, we claim that it is not just some sample data: it will allow a suitably placed observer to learn something about the real world (Dretske 1981), at the time the information was constructed. This field of thinking has its own thought-experiments, for example, a bear-track in the woods contains the information that a bear passed that way whether or not anyone ever notices the spoor, and anyone suitably placed to notice it can learn this content.
The raw data from a survey, or a sheaf of newspaper cuttings, may well contain a lot of information that can be drawn to our attention by a suitably placed researcher. Unlike the bear-track in the woods, however, the survey was collected, and the newspaper articles written, by humans who have well-known tendencies to misperception, mistake, misinformation… So while the Internet doubtless contains a lot of useful information, I know it also contains much that is erroneous, ill-informed, and misleading. In research we don’t accept any of it uncritically. We try to stick to good sources of information, we try for honesty in our evidence gathering, and we try to take care in our conclusions, in our mission to increase the stock of knowledge in our academic discipline.
In my June 20 contribution to this blog I rather incautiously mentioned the pursuit of the truth about (the laws of) the physical world, and I hinted that I felt this was not the business of science, or research for that matter. Truth is a great concept and a noble ideal, but as I mentioned in the April 18 contribution, the more truthful we make any statement, the less clear it is, and the narrower the scope of its application. In a way, the only absolutely truthful statements are the formal tautologies of pure mathematics, the necessary truths that depend on nothing, and so add nothing to our knowledge. Conversely, any statement that is not necessarily true is merely (by definition) contingently true. It might be true, that is, if (um..) everything were really as it says. There are few philosophers, and even fewer scientists, who feel that this sort of circular discussion is worth while.
And yet, precisely this kind of argument has fascinated people for millennia. Descartes’ necessary truth “cogito, ergo sum” (1641) was the start of his argument to prove the existence of God and the immortality of the soul. A very similar exercise by Bernard Lonergan (1904-1984) achieved a wide currency in 1956. Now the reader should always smell a rat if someone claims to prove a contingent truth from a necessary one. The Australian philosopher David Stove (1927-1994) catalogued a large number of such arguments from Plato to Habermas: explaining that since they want to make everyone accept their opinion, it is a good trick to make it appear merely a logical deduction from a necessary truth. The trick can be made to work, as he explains, with the support of some impressive but contradictory concept, in the same way that division by zero can be used to create a convincing-looking proof. Usually though, such philosophers are vainly trying to use logic to establish some belief which predates their attempt and will outlast their failure.
To return to computing: every piece of information, according to all of us who follow Floridi and Dretske, contains its very own claim of contingent truth. By labelling it as information, we claim that it is not just some sample data: it will allow a suitably placed observer to learn something about the real world (Dretske 1981), at the time the information was constructed. This field of thinking has its own thought-experiments, for example, a bear-track in the woods contains the information that a bear passed that way whether or not anyone ever notices the spoor, and anyone suitably placed to notice it can learn this content.
The raw data from a survey, or a sheaf of newspaper cuttings, may well contain a lot of information that can be drawn to our attention by a suitably placed researcher. Unlike the bear-track in the woods, however, the survey was collected, and the newspaper articles written, by humans who have well-known tendencies to misperception, mistake, misinformation… So while the Internet doubtless contains a lot of useful information, I know it also contains much that is erroneous, ill-informed, and misleading. In research we don’t accept any of it uncritically. We try to stick to good sources of information, we try for honesty in our evidence gathering, and we try to take care in our conclusions, in our mission to increase the stock of knowledge in our academic discipline.
Saturday, June 27, 2009
Supervision
Whose research is it? Yours. The basic idea will be something that the supervisor is an expert in (or you need another supervisor), but since you are making a new contribution to knowledge, at the end of the process you will be the world expert in your subject.
Your supervisor plays hugely important roles throughout the process though. At the start, they will help you with the relevant literature and established approaches. As the research progresses, they will help with methodology, with planning the research, and helping you phrase your research questions. Once you have parts of your thesis in draft, they will provide an invaluable critique of the flow of argument, and the construction of your thesis as a piece of rationally-argued writing. Your supervisor will also play a crucial role in selecting your external examiners, and being your supporter and eyes and ears during the viva.
Above all, throughout the process, they are following your journey, engaging in the discussions, playing the part of reader of your thesis and papers, reacting in the ways that your audience and your examiners might to the parts of your work that are new and surprising, so that you can fine-tune your arguments and make sure there are no loose ends.
The relationship between student and supervisor can sometimes be stormy - it is always a two way process, and a second supervisor can sometimes play a useful role in getting things back on track. It can and should be inspiring.
Your supervisor plays hugely important roles throughout the process though. At the start, they will help you with the relevant literature and established approaches. As the research progresses, they will help with methodology, with planning the research, and helping you phrase your research questions. Once you have parts of your thesis in draft, they will provide an invaluable critique of the flow of argument, and the construction of your thesis as a piece of rationally-argued writing. Your supervisor will also play a crucial role in selecting your external examiners, and being your supporter and eyes and ears during the viva.
Above all, throughout the process, they are following your journey, engaging in the discussions, playing the part of reader of your thesis and papers, reacting in the ways that your audience and your examiners might to the parts of your work that are new and surprising, so that you can fine-tune your arguments and make sure there are no loose ends.
The relationship between student and supervisor can sometimes be stormy - it is always a two way process, and a second supervisor can sometimes play a useful role in getting things back on track. It can and should be inspiring.
Saturday, June 20, 2009
All models are wrong
.. but some models are useful (George Box et al, 2009, p.61). What makes a model useful? Some theories of science have made grand descriptions in terms of prediction, explanation etc, but it really comes down to a consensus. Today (June 20th) it is reported that the British Government has decided that the spelling rule “I before e except after c” should no longer be taught in schools because the large number of exceptions made it useless. Such a rule is of course one of observation rather than a law of nature, but on close inspection it is easy to find hidden qualifications to any law you care to mention.
Until very recently, many in the scientific community used to imagine that they were discovering the truth about how the physical world works. Whewell (1833, p.256) quotes Lagrange’s opinion “that Newton was fortunate in having the system of the world for his problem, since its theory could be discovered once only”. Now a lake can be discovered only once, but systems are merely constructed, and many refinements and re-interpretations will be possible. Twentieth century physics revealed unimaginable strangeness, needing many alternative and conflicting models to apply to quantum mechanics, diffraction, cosmology, etc., and there was some useful criticism of old notions such as “final causes” (basically, boundary conditions at infinity).
Many researchers in the late nineteenth and early twentieth century searched only for natural laws expressible in terms of differential equations. Since this search followed so closely after the development of the calculus it appears with hindsight that these men with a new hammer suddenly saw nails everywhere.
The same hindsight opens our eyes to the serious untruths in their “natural laws”: on close inspection a natural law does not actually apply everywhere, but only (um..) where it applies (e.g. in the absence of discontinuities, in a neighbourhood of the origin). To be fair, talk of truth or laws was mostly a habit of speech: the models described in these laws are useful in telling us what to expect in the sort of situation for which the model was designed. In other situations or on close inspection we might need a different or more refined model.
As with final causes, or the ether, models can be useful even when they conflict with other models (seem counter-intuitive) or don’t fit with current ideas of causality. For example, classical field theory remains useful, even though we know that action at a distance is impossible, and that there is a better model based on radiation. Non-existent lakes will eventually be removed from atlases, but models will continue as long as somebody finds them useful.
References:
Box, G. E. P; Luceo, A.; Paniagua-quinones, M. d. C. (2009): Statistical Control by Monitoring and Feedback Adjustment, 2nd ed. (Wiley) ISBN 0470148322
Whewell, W. (1833) Astronomy and General Physics: Considered with Reference to Natural Theology
Until very recently, many in the scientific community used to imagine that they were discovering the truth about how the physical world works. Whewell (1833, p.256) quotes Lagrange’s opinion “that Newton was fortunate in having the system of the world for his problem, since its theory could be discovered once only”. Now a lake can be discovered only once, but systems are merely constructed, and many refinements and re-interpretations will be possible. Twentieth century physics revealed unimaginable strangeness, needing many alternative and conflicting models to apply to quantum mechanics, diffraction, cosmology, etc., and there was some useful criticism of old notions such as “final causes” (basically, boundary conditions at infinity).
Many researchers in the late nineteenth and early twentieth century searched only for natural laws expressible in terms of differential equations. Since this search followed so closely after the development of the calculus it appears with hindsight that these men with a new hammer suddenly saw nails everywhere.
The same hindsight opens our eyes to the serious untruths in their “natural laws”: on close inspection a natural law does not actually apply everywhere, but only (um..) where it applies (e.g. in the absence of discontinuities, in a neighbourhood of the origin). To be fair, talk of truth or laws was mostly a habit of speech: the models described in these laws are useful in telling us what to expect in the sort of situation for which the model was designed. In other situations or on close inspection we might need a different or more refined model.
As with final causes, or the ether, models can be useful even when they conflict with other models (seem counter-intuitive) or don’t fit with current ideas of causality. For example, classical field theory remains useful, even though we know that action at a distance is impossible, and that there is a better model based on radiation. Non-existent lakes will eventually be removed from atlases, but models will continue as long as somebody finds them useful.
References:
Box, G. E. P; Luceo, A.; Paniagua-quinones, M. d. C. (2009): Statistical Control by Monitoring and Feedback Adjustment, 2nd ed. (Wiley) ISBN 0470148322
Whewell, W. (1833) Astronomy and General Physics: Considered with Reference to Natural Theology
Saturday, June 13, 2009
Evaluation and testing
If your contribution to knowledge is a better way of doing something, coming up with the idea (and maybe implementing it somehow) is only half the battle. The real work will come with evaluation, and this will need a methodology all on its own.
Most new algorithms are tested against data sets found in the literature and things are interesting if in some sense your idea performs better than its predecessors in these tests. But there are often problems with this approach – and I have occasionally seen a conspiracy of silence where anyone can see that the comparison is not entirely fair. After all, if your new algorithm is suited to a particular class of problem not previously addressed, then why test in some previous or different scenario? There are difficulties in using data from a different problem scenario, or comparing with an algorithm that was actually tackling a different issue. This sort of test is rarely more than the researcher’s equivalent of “Hello world”. The conspiracy of silence arises because this test is standard in the literature, and is used for convenience even though everyone knows that the data is unrealistically simplistic, or has been cleaned to remove any real-world difficulties.
On the other hand, artificial data, designed to exhibit the sort of issue that your idea helps to solve, has a value. It is dealing with the sort of hypothesis that starts “Problems reported with this aircraft control system may be associated with …” and your investigation is as much about exploring some peculiar feature that might occasionally occur in the data, and exploring responses to such a feature from the existing algorithms and yours, to see if the hypothesised feature was the source of the actual difficulty.
If your contribution arises from some solving a real-world problem, it will probably need a lot more work to collect real-world data and draw real-world conclusions. Space and time considerations limit most PhD theses, and all conference papers, to artificial, toy examples. Maybe though, some preliminary data can be analysed within the PhD (and lead to a job with the company with the problem so you can work on the real data) and the evaluation part can be beefed up with actual comments about the value of the contribution from those more familiar with the real-world problem.
In addition, the implementation of the idea in your PhD will probably have the nature of a prototype, which will need to be re-implemented within a real-world control system. Again, space and time considerations make it unlikely that real production software will be used in your PhD or any publication resulting from it. But actual adoption of the approach within the industrial process will count as a complete proof of the value of your contribution, and any progress in this direction should go in your thesis.
Most new algorithms are tested against data sets found in the literature and things are interesting if in some sense your idea performs better than its predecessors in these tests. But there are often problems with this approach – and I have occasionally seen a conspiracy of silence where anyone can see that the comparison is not entirely fair. After all, if your new algorithm is suited to a particular class of problem not previously addressed, then why test in some previous or different scenario? There are difficulties in using data from a different problem scenario, or comparing with an algorithm that was actually tackling a different issue. This sort of test is rarely more than the researcher’s equivalent of “Hello world”. The conspiracy of silence arises because this test is standard in the literature, and is used for convenience even though everyone knows that the data is unrealistically simplistic, or has been cleaned to remove any real-world difficulties.
On the other hand, artificial data, designed to exhibit the sort of issue that your idea helps to solve, has a value. It is dealing with the sort of hypothesis that starts “Problems reported with this aircraft control system may be associated with …” and your investigation is as much about exploring some peculiar feature that might occasionally occur in the data, and exploring responses to such a feature from the existing algorithms and yours, to see if the hypothesised feature was the source of the actual difficulty.
If your contribution arises from some solving a real-world problem, it will probably need a lot more work to collect real-world data and draw real-world conclusions. Space and time considerations limit most PhD theses, and all conference papers, to artificial, toy examples. Maybe though, some preliminary data can be analysed within the PhD (and lead to a job with the company with the problem so you can work on the real data) and the evaluation part can be beefed up with actual comments about the value of the contribution from those more familiar with the real-world problem.
In addition, the implementation of the idea in your PhD will probably have the nature of a prototype, which will need to be re-implemented within a real-world control system. Again, space and time considerations make it unlikely that real production software will be used in your PhD or any publication resulting from it. But actual adoption of the approach within the industrial process will count as a complete proof of the value of your contribution, and any progress in this direction should go in your thesis.
Saturday, June 6, 2009
Ethics and bad research
Like “health and safety”, the phrase “research ethics” tends to elicit weary groans from many researchers. A full discussion is obviously out of place in this blog, but it seems obvious that research should not do actual harm without a very convincing argument. What I would like to focus on is whether bad research is ever ethical.
By bad research I mean research that is poorly thought out, where data cannot reliably support the sort of investigation for which they were collected. People have used up time, and costs incurred, for no benefit. In my view such research is always unethical, since its value (roughly zero) does not justify the trouble it has taken. It may cause actual harm, possibly even to the whole process of research, if enough people find it ridiculous. Future funding, or the cooperation of potential subjects, may be affected if research does not seem to be useful.
You therefore need to explain why your research really is useful, and why your subjects have to answer a long list of strange-looking questions. This explanation is for when you approach potential subjects, supervisors and sources of funding. You need to be open about what your research is about and what its expected benefits are. You must not use any deception in your approach to any of these people. Not can you say (yet) what the conclusions will be. Sometimes (rarely?) it will not be possible without compromising the research outcomes to tell your subjects what the hypothesis is, but you must be able to explain the expected area of benefit of the research and why they have been approached. Also, you should never collect data without discussing these aspects first.
An anecdote may help explain this point. Suppose you are at a management training course and you are given a set of objectives to prioritise. It is late in the day, and they all look important so you just take the first six and make up some spurious reasons for your choice. You could well be irritated if these priorities are fed back to senior management in your company – in two ways. Because your careless reasoning may be subjected to more scrutiny tan you would like, and this reflects badly on you. But more importantly, you fear that these may be the wrong priorities, and if you had known they would be used, you would have taken more care over them. Because of the careless way the data has been collected you do not know if your performance will be unfairly judged, or if the organisation will now change its behaviour as a result of bad data.
The selection of sources of data, whether from human subjects or more generally, requires the greatest care, and has been discussed in an earlier entry in this blog, as the chosen pattern will have a crucial bearing on the scope of validity of your conclusions. What data you collect, and how, will limit its interpretation, and this too has been discussed in the entry on research methodology. You probably won’t need to get ethical clearance unless your research involves living subjects, but the application you make before you start will provide a concise overview of the plan for your research and how the conclusions will be drawn. Even if you don’t need ethical clearance, you should protect your research by thinking these things out.
By bad research I mean research that is poorly thought out, where data cannot reliably support the sort of investigation for which they were collected. People have used up time, and costs incurred, for no benefit. In my view such research is always unethical, since its value (roughly zero) does not justify the trouble it has taken. It may cause actual harm, possibly even to the whole process of research, if enough people find it ridiculous. Future funding, or the cooperation of potential subjects, may be affected if research does not seem to be useful.
You therefore need to explain why your research really is useful, and why your subjects have to answer a long list of strange-looking questions. This explanation is for when you approach potential subjects, supervisors and sources of funding. You need to be open about what your research is about and what its expected benefits are. You must not use any deception in your approach to any of these people. Not can you say (yet) what the conclusions will be. Sometimes (rarely?) it will not be possible without compromising the research outcomes to tell your subjects what the hypothesis is, but you must be able to explain the expected area of benefit of the research and why they have been approached. Also, you should never collect data without discussing these aspects first.
An anecdote may help explain this point. Suppose you are at a management training course and you are given a set of objectives to prioritise. It is late in the day, and they all look important so you just take the first six and make up some spurious reasons for your choice. You could well be irritated if these priorities are fed back to senior management in your company – in two ways. Because your careless reasoning may be subjected to more scrutiny tan you would like, and this reflects badly on you. But more importantly, you fear that these may be the wrong priorities, and if you had known they would be used, you would have taken more care over them. Because of the careless way the data has been collected you do not know if your performance will be unfairly judged, or if the organisation will now change its behaviour as a result of bad data.
The selection of sources of data, whether from human subjects or more generally, requires the greatest care, and has been discussed in an earlier entry in this blog, as the chosen pattern will have a crucial bearing on the scope of validity of your conclusions. What data you collect, and how, will limit its interpretation, and this too has been discussed in the entry on research methodology. You probably won’t need to get ethical clearance unless your research involves living subjects, but the application you make before you start will provide a concise overview of the plan for your research and how the conclusions will be drawn. Even if you don’t need ethical clearance, you should protect your research by thinking these things out.
Subscribe to:
Posts (Atom)
