ISI Durban

[Chris Lloyd]

The ISI 2009 conference was held lst week in Durban, South Africa. I quite enjoyed the previous one in Lisbon and was invited to this one so I thought I would go back for seconds. Below the fold are some impressions (considered an uncontroversial as ever) on the academic and non-academic aspects of the meeting.

First, I will admit that I am cranky with the organizers for listing my special session wrongly on the timetable and then not advertising the reschedule, or the location. I couldn’t even find the room myself. Consequently there was an audience of the six which included the four speakers. Thanks are due to Paul Kabaila for making up half of the non-presenting audience.


The more serious gripe with the conference was personal safety. In Durban you cannot actually walk the streets. Anyone who might have money (which includes all whites) is*fairly likely to get mugged and robbed. The hotels were no more than 400 meters from the conference venue, the roads were open and wide, the sun was shining. But the conference organizers and hotel managers insisted that we could not walk it, even at 9 in the morning and even in large groups. After dark, don’t even think about it. Unaccountably, there is very little visible police presence. It is no way to live and, I would have thought, death to the South African economy.


Probably the non-academic highlight of the conference was the game park tour. You are actually safer amongst the lions and rhinos than you are walking the predatory streets of Durban! It was interesting to see the animals interacting together in a quasi natural environment rather than being on display in separate cages. For instance, the smaller animals all hang around the giraffes and zebras who are the best as seeing and smelling predators respectively. Free riding is rife in nature.


Turning to the substance of the meeting, I don’t know what your experiences are, but I am finding more and more that, at large conferences at least, most of the talks are simply awful. There are entire sessions of people reporting uninteresting applied work,*rediscoveries of*well known results, papers that are technically incomprehensible to anyone, speakers reading directly from slides in unintelligible English and an increasing incidence of “no shows”.


Unlike some other disciplines, when you send your abstract to a statistics conference and receive the email saying “your paper has been accepted”, we all know that nobody has actually read the paper*. There seems to be no quality control at all as far as I can see. I guess the organizers just want as many paying participants as possible to cover their costs.


But enough of the negative! There were a couple of sessions I enjoyed where I learned something new and fun.


The first concerned distributions theory, specifically distributions that are defined more easily through their quantile function (i.e. inverse cdf) than their cdf or pdf. Examples are the Davies distribution and the generalized lambda distribution. Ultimately, you can always get the cdf numerically and just use standard methods but the alternative quantile view leads to new ideas and techniques. First, there are a bunch of moments from integrating the quantile function with respect to Legendre polynomials - called Legendre or L*moments. The estimators are linear combinations fo order statistics. If the first Legendre moment exists then they all exist. The distribution with standard zero Legendre moments it the uniform dsitribution, not the normal. These exotic oments generate alternative method of moments estimators. Second, quantile functions combine simply. Location and scale transforms of the quantile function transform the location and scale of the variable. Also a linear combination of quantile functions is another quantile function but it does not correspond to a mixture. Only the odd moments are affected. This makes it easy to augmenting standard distributions into larger families. Anyway, while there was nothing earth shattering and probably nothing presented that was unknown, it was a rather new and fresh – to me at least.


The second session I enjoyed was on adversarial risk analysis by which they mean quantitative strategies against terrorists. David Banks, the organizer and a main speaker of the session began by personally introducing himself to each audience member (there were only about 10 of us), thanking us for coming and asking us what our interests and background were. I use this angle for cutting the ice with a new MBA class but I have never seen it used at a statistical conference before.


The session was about using decision trees where there are two players with different utilities on the possible outcomes and where the utilities and probabilities for the terrorist are unknown. Standard game theory won’t work because the utilities are not the same for the two players nor known to each other. Standard decision analysis also fails in practice because the unknown actions of the adversary are not meaningfully dealt with using probability.


The proposed approach was put yourself in the terrorist shoes, to elicit a distribution for the terrorist’s assessment of your utilities and probabilities, and use this to generate a distribution for his probabilities. This can also be pushed one step further back, where you*make estimates of*what the terrorist will likely assume about your probabilities and utilities. Pushing the subjective probabilities deeper can actually lead to better robustness, kind of like Bayesian hyper priors having less influence than direct priors. So it is a combination of game theory, subjective probability elicitation and decision trees. Most of the details are in this recent JASA paper.




* Note to self. I might test this assertion by sending an abstract containing deliberate nonsense and see if it is accepted – like a nerdy statistical version of Angry Penguins.



Get More from the original blog...