Those Pesky Humans: Urban Planning and its Discontents

Article first published as Those Pesky Humans: Urban Planning and its Discontents on Blogcritics.

Greg Lindsay writes in the New York Times that Pegasus Holdings, a technology company based in Washington, DC, is building a "medium sized town" on 20 square miles of New Mexico desert. The town, dubbed the "Center for Innovation, Testing, and Evaluation" (mark it on the map!), will contain infrastructure adequate to support a population of 35,000, but will be home only to a handful of engineers and other geeks from Pegasus, who plan to use it as a laboratory to build future "smart cities", where power grids, traffic, security and surveillance systems are monitored and controlled by computer.

On the face of it, "smart cities" sound like a good idea (better than, say, "dumb cities"). The idea is, in outline, simple enough: a) install sensors to get information about how people move about and interact in cities, then b) feed this data to computers develop complex models of human behavior, generating policies that make things work better, more efficiently. To take an obvious example, who wouldn't want traffic lights optimized to increase vehicle throughput? Or pedestrian pathways that make two-way foot traffic flow more smoothly? Makes sense, right?

Yet, as Lindsay points out, these seemingly innocuous examples paper over a broader project that has repeatedly been exposed as folly, that of trying to simulate the behavior of people in cities using abstractions like computer models, rather than by gaining an understanding of what people living in cities care about, and find valuable. These qualitative, subject elements are typically what determine what makes a great city "great", smart by computer modeling standards or not.

It would seem obvious and necessary to account for this "human-factor" when constructing quantitative models for smart city projects like Pegasus' (after all, we're talking about humans), only, as is so often the case, the computer geeks view "qualitative" features of a city as the very thing that needs to be analyzed quantitatively, and replaced. As Rober H. Brumly, managing director and co-founder of Pegasus pronounced, "We think that sensor development has gotten to the point now where you can replicated human behavior".

And so Brumly and the Pegasus visionaries, in this latest round of "machine versus man", continue the tradition of remaining seemingly ignorant of the manifest lessons of over-thinking urban planning going back decades, at least to the publication of the seminal "The Death and Life of Great American Cities" by flesh and blood New Yorker Jane Jacobs. Jacobs repeatedly documented how best laid urban plans would lead to frustration and a sense of alienation in the neighborhoods of New York City. For example, urban planners who attempted a gentrification project in a NYC slum decided that planting strips of grass outside tenements would have a salubrious effect. But alas, the pesky human tenants saw the grass strips as ridiculous, ill-placed, and insulting. It had the opposite effect, in other words, which could have been "predicted" had only the urban planners taken the time to understand the neighborhood, and get to know the tastes and circumstances of its inhabitants.

And there are more nefarious examples, like the 1968 RAND project to reduce fire response times in NYC, resulting in an estimated 60,000 fires in impoverished sections of New York, as "faulty data and flawed assumptions" triggered the replacement of fire stations in Brooklyn, Queens, and the Bronx with smaller ones. The coup de grace here was the politicization of the supposedly "scientific" project, where clever RAND officials, realizing that rich folk in well-to-do neighborhoods would not tolerate the effects of "efficiency" using their (flawed) simulations, placed such neighborhoods outside the scope of the project.

And on and on the story goes. Unintended consequences are simply part and parcel of the development of causal or predictive models using quantitative data gleaned from messy, complex systems. The real folly, however, in the Pegasus project and so many others like it, is not in the (basically correct) idea that quantitative analysis can provide useful information when devising strategies, for urban planning or otherwise, but that the human element can therefore be eliminated. That latter claim does not follow, and taking it too seriously will almost certainly guarantee that among the lessons we learn from the "Center for Innovation, Testing, and Evaluation", one of the most important is likely to be that innovation, testing, and evaluation is not enough.

I Eat Yogurt, Therefore I Am

In the WSJ today, Jonah Lehrer, who dropped out of his Ph.D. program in neuroscience to make millions writing Barnes and Noble science books like "Proust Was a Neuroscientist", and "How We Decide", wrote a piece in the review section about eating yogurt and its connection to the mind-body problem. The basic idea is that yogurt makes you less anxious, because it contains probiotics, which contain GABA, a neurotransmitter that limits the effects of neurons. This is all true enough, I'm sure, just as its true that eating simple carbohydrates gives one a feeling of energy followed by a "crash". It's no mystery that the types of foods we eat affect how we feel. But it's quite a leap from this sensible factoid to conclusions about the nature of the mind--if it's distinct from the brain, or more generally our physical bodies. In fact Lehrer glosses over the pivotal conceptual conundrum, that all the gastronomic observations he or anyone else adduces in favor of theories about the nature of mind are consistent with theories that correlate mind and body, as well as those that identify them. C'mon Jonah, you surely most know this. Was it that hard to find something to say?

The Wisdom of Crowds

I had an interesting discussion with one of my (many) liberal friends recently, and there's a couple of points that came out of our prolonged exploration of every idea we could think of in the span of an evening. The first point worth sharing, I think, is as follows. The barnes and noble science section is packed with books explaining that we can't predict the future in systems that aren't governed by so-called normal distributions (though we think we can, a kind of persistent overconfidence in our epistemic abilities). The point here is that there's all of this complexity in everyday life and social life (think economics), and we're under the illusion that having a Ph.D. in economics and pointing to charts renders all of this moot. I'm talking about Taleb's "Black Swan", books like "The Drunkards Walk", everything that Malcom Gladwell has every written, et cetera. So the point is, it turns out that huge parts of the world--and interestingly, ordinary parts of the world, like society and culture, politics, economics--are effectively black boxes with respect to prediction. We really just don't know what tomorrow will bring. This has implications--huge implications--for the role of government or in general the role of experts in advising the rest of us on what courses of action should be undertaken. Much of this advice should properly be seen as speculation (there's even research that suggests that experts are actually worse at predicting outcomes in complex "human" systems than non-experts).

So that's point one. The other point is the "wisdom of crowds" notion, another concept that accounts for dozens of books in the BN science section (you know, where the smart-people-wanna-bes congregate). So this idea that many problems are solved by aggregating viewpoints, explored in books like, ah hum, The Wisdom of Crowds, Infotopia, Jeff Howe's Crowdsourcing, et cetera, sugggests that having some expert decide things can be really stupid. In fact it turns out that groups--crowds--can often arrive at more optimal solutions to problems then even some one person who is educated and expert on solving problems of that type. It turns out, for instance, that allowing people to "bet" on some future outcome often produces the best prediction of that outcome. Asking an expert to figure out the outcome would result in a worse prediction. Just ordinary Joes (and of course experts too), if there's enough of them, throwing down their money on which horse will win, or which companies to buy stock in, or which presidential candidate will win the election, often produces a better prediction on aggregate than the most expert horse person, or stock picker, or election pundit.

So this cutting edge research makes everyone feel all enlightened and egalitarian and up to date with the latest tidbits of "didn't you know?" science. Only thing is, this is the best, greatest empirical argument for free markets ever. Let ordinary people figure out what to buy, where to shop, how the economy should go, from the ground up, so to speak. Science suggests that this libertarian technique often results in more optimal solutions then central planners sitting in government offices. This really amuses me, because the folks feeling all educated reading wisdom of the crowds literature-- folks interested in social networking technology, reading Marx, sniffling about how republicans are idiots, eating tofu-- yes these folks are in fact reading powerful arguments for individual liberty, limited government, the wisdom in crowds, not government planners trying to engineer the Good Society for everyone else. (And they seem not to know it. Ha!) The latter just isn't optimal, if you believe the latest Gladwellesque arguments coming out of research on decision making.

Which brings me to my little wrap up. When I'm not sitting in cafes and I'm actually doing serious work, I'm reading Hayek, the Nobel economist who is widely credited as an intellectual precursor to libertarianism, particularly with regard to government involvement in the economy. Hayek, that Ph.D. egghead himself, who nonetheless argued (in my view persuasively) in the 1950s that because no one person can possibly know everything, the best strategy for a society is to vest more and more power in individuals. Hence individual liberty is a strategy that is most likely, over time, to result in more optimal solutions. Makes sense. The wisdom in crowds.

So the point is, again, that all of these insights emerging from the latest research in the social sciences are pointing back to non-government-controlled solutions to our most pressing problems. It's pointing to a model where government's most important job is to structure society in such a way that its citizens--all of us--can choose how best to live, what to buy, who to give money to, how to use our own money, and on and on. The aggregation of all of these individual voices makes things work better (right comrades?). But of course people who are free to choose and to decide much of their lives on their own means that disparities in wealth and natural talents will result in disparities in society. And that bothers all of my liberal friends. "Can't we just control people and liberate them too?" perhaps I heard (didn't they tell us in school? you have to control people to try to make them equal, since we're not naturally that way). But no I think unfortunately, on the aggregate, there's a better and worse way to do things. And on the aggregate, their will always be winners and losers in the crowd.

Obama Jumps Shark on Mars

President Obama re-iterated his commitment to the space program today, announcing that he expects Earthlings to reach Mars within his lifetime. Speaking at the Kennedy Space Center in Cape Canaveral, Florida, Obama sought to dispel any skepticism of his support for NASA, and added that not just astronauts but ordinary Americans, "like tea partiers", could soon be shipped to Mars as part of his new space exploration program.

Gotcha. Really, though. I'm ambivalent about getting crabby about this latest salvo from Big Idea Obama (because in general I like the space program, and I like Big Ideas), but nonetheless it seems that we would all do well to observe some hard to define yet common sense interval between Big Ideas, wouldn't we? From historical revamping of health care, to reduce the nukes (an issue we've become accustomed to ignoring, true), to humans on mars, what's next? They're all darn worthy proposals for discussion, debate, analysis, action. But it'd feel more grounded, more deliberative and thought through and, well, presidential, if we had some time to close out discussion on one Big Idea--it's benefits, consequences, downsides--before launching into the next. But, alas, fat chance for quiet time with Big Ideas, it seems. The presidency these days is all 'bout hittin' the turbo thrusters (think: the Iraq War and former President Bush).

Well, anyway. If my political compass is in working order, at least we won't hear much more than a whimper about "cap and trade" or "global warming" from the White House. The dire warnings of impending disaster emanating from Gore and other catastrophe types have quite clearly fallen out of favor these days (ever notice how, once the hysteria goes away, you can't get people to defend the original thesis anymore? They defend milk toast versions of it, the kind of discussion that you wanted to have back when everyone was screeching, but they thought too moderate. Ask a Global Warming proponent to talk about Global Warming now, and you'll get discussions about pollution, and reasonable-sounding stuff about how we should get rid of carbon fuels anyway, et cetera. What happened to Catastrophe!? It all just kind of gets forgotten, as if there really wasn't a hysteria in the first place, and we all just move on in a state of collective amnesia about our cries of imminent doom just two or three years prior. Ugh.).

Anywho. Mars! Cool. Next?

Systems Accidents

Laurence Gonzales, in his best-selling Deep Survival, discusses a class of accidents that happen in complex, real-world systems, called systems accidents. The term was first introduced by John Perrow in his now famous work on accidents, Normal Accidents, published in 1984. Perrow argued that in certain kinds of systems, catastrophic accidents, while rare, are inevitable. What's worse, and what made Perrow's work so controversial, attempts to avoid such accidents only make them more inevitable, because the safety measures put in place to expunge them also make the systems where they occur more complex. It is precisely the complexity of such systems which make them prone, one way or the other, to rare but catastrophic failure.

Gonzalez (citing Perrow) points to two features of complex systems that make them prone to systems accidents. One, the systems must support "unintended complex interactions among components and forces". Two, they must be "tightly coupled", or constructed in such a way that forces, even if initially small, can magnify as they propagate from component to component. In systems with such features, the result of forces acting on or in the system cannot be exactly known: seemingly insignificant events can chain together in unexpected ways, leading to unintended results (feature one), and some of these interactions can have global effects: the system itself can be "blown apart" by ostensibly innocuous forces that, by cause and effect, release the energy in the whole system destructively (feature two). Given these features, systems accidents will occur; they are a natural result of the complexity of such systems.

Gonzalez mentions a few examples of Perrow's complex systems, like the modern airline, a "large mass containing explosive fuel, flying at high speeds, and operating along a fine boundary between stability and instability". Something seemingly insignificant happens (say, the onboard toilet malfunctions), and it leads to a catastrophic failure, one that is nearly impossible to predict in advance and, as Perrow argued, cannot be eliminated in kind, because the safety components one might introduce to solve some problems will themselves create complexities that lead to other problems that won't be understood until an accident occurs, too. (The toilet leading to a catastrophic crash is a real example.)

Gonzalez connects Perrow's work on systems accidents with branches of mathematics and science concerned with so-called nonlinear systems, like chaos theory. Chaos theory has entered into the popular or "pop" science vocabulary (it was mentioned in the movie Jurassic Park, for instance) as a catch-all phrase for systems with states that are difficult or impossible to predict because they have a "sensitive dependence on initial conditions", as science writer James Gleick put it in his 1987 book Chaos. Turbulence is a classic example of a chaotic system: applying physics equations to predict where a cubic inch of water in a turbulent stream will be in, say, 10 seconds is effectively impossible, because the initial conditions of the cube of water--including all of the factors that can act on it--have to be specified so exactly, and again and again as the cube propagates in the turbulent stream, that no amount of computation can possibly determine its later position. True, the prediction is possible in principle, since the laws governing dynamic systems like flowing water are known, but not in practice, because the calculations we need for prediction require such an exactitude and are so complex that not even fantastically powerful computers (say, composed of all of the matter in the universe) could compute predictions of the position of the cube after some (relatively small) window of time. It's effectively impossible to predict the downstream position of the cube of water, in other words. This realization, that chaotic systems give rise to unpredictable behavior, led to notions about unpredictability like the "butterfly effect", where it's quipped that the air displaced by a butterfly's wings in Tokyo might, a month later, create a hurricane in the Gulf of Mexico.

Chaos theory applies to Perrow's complex systems because, just like with the butterfly effect, such systems have a sensitive dependence on the initial conditions that might lead, ultimately, to systems accidents. No one thought about the toilet malfunction beginning a chain that led to the crash; similarly, the sensitivity of O-rings (remember them?) to cold weather was known before the Challenger explosion, but no one put together that the particular O-rings installed, along with the exact temperature in Cape Canaveral the morning of the fateful 1986 launch, would by cause and effect result in the explosion. But it happened, and if Perrow is right, such systems accidents will always happen, one way or the other. Some types of systems accidents will be analyzed and fixed, but others will emerge, and while they will remain rare, they will remain inevitable, too.

What's interesting about Gonzalez's treatment of systems accidents via Perrow is his application of them to human hobbies and activities. He tells stories of experienced mountain climbers who are killed precisely because they applied their experience to situations, and put in place safety measures based on what worked in past attempts. By and large, the safety measures climbers employ do work: teams of climbers roped together help reduce the likelihood that a particular climber will fall, because when someone slips, the other climbers can "self-arrest", which means they plant their ice axes in the snow and help break the fall of the climber who slipped. But sometimes, they don't.

They key to understanding Gonzalez's examples is to get his distinction between general trends and specific situations. Safety systems address general risk: mountain climbing is in general more safe today than, say, a hundred years ago. Like a seatbelt worn in a car, the safety procedures climbers employ on mountains function to lessen the number of accidents and the damaging consequences of them when they do occur. But, as Gonzalez points out, sometimes the safety system itself causes or exacerbates an accident in a specific (unpredictable) situation, just as Perrow argued it will. The safety system introduces massive amounts of energy into systems of climbers, making them tightly coupled, and it introduces the possibility that some situations will lead, by cause and effect, to the "blowing apart" of the constructed system, as all of that energy put into the system (for safety), suddenly becomes the energy that magnifies the accident. As Gonzalez describes so well in Deep Survival, this is exactly what happened to a group of experienced climbers who self-arrested on Mount Hood to save a member of their team from falling. Several people died in the accident; had the safety measures not been used, one person would have fallen, and though this is of course tragic, it's not worse than what did happen, which became one of the worst mountain climbing disasters ever.

The point of Deep Survival is not that we should throw out safety measures, or give up on making them better. The point is that systems create complexities--whether in nature, or by our own design--that limit prediction and make deterministic solutions impossible. Chaos and complexity gaurantee that life will remain messy, incomplete, and in large part unpredictable. They gaurantee that, try as we might, we'll never, as Dostoevsky once remarked, erect a "crystal palace of reason", where everything is in place, and understood, and predictable. Systems accidents are here to stay, and though our best efforts help reduce the inherent risk of living and acting in the world, they can't eliminate it, and in many cases they succeed in making the risk that remains even graver, and the accidents we occasionally do have even more catastrophic.

Sam Harris Talk

An interesting talk by Sam Harris, of "The End of Faith" fame, at TED. This is worth listening to, about 18 minutes. His idea is that there are scientific facts that tell us a lot about what we should value, which begins to bridge the gap between science and morality, ultimately making it possible to have a "science of morality".

Not Just Madness

A scene in the eminently watchable Michael Clayton shows a bedroom wall adorned with a message from lawyer-gone-mental-patient Arthur Edens (played by Tom Wilkinson): "Make Believe It's Not Just Madness".

What is it supposed to mean? I won't get into the movie, but will offer the following interpretations:

(1) It's not only madness, it's also something else, unspecified. It could be madness and indigestion. Or madness and foot rot. Pretend it's so.
(2) It's not only madness, it's a mix of madness with other explanatory factors, like jealosy, or shame, or guilt. Pretend it's this larger web of humanizing factors. Not just madness.
(3) It's madness, but it's not just. He had no right to go mad. So, make believe it's madness point blank, without just cause.

Is it ((not just) madness) or (not (just madness))?

The Neanderthal Project

The NYT recently reported that DNA sequencing of the Wooly Mammoth genome is now possible, using two fossilized hair samples, recovered from mammoths that died 20,000 and 60,000 years ago. NYT reports that scientists are now discussing how to modify DNA in the mammoth's closest living relative, the African elephant, so that it resembles the wooly mammoth. The elephant genome, according to Stephan C. Schuster and Webb Miller at Penn State, will need to be modified at about 400,000 places to make it resemble its hairer cousin. As the thinking goes, once modified at these locations, the elephant genome will be, effectively, a woolly mammoth genome, which can then be brought to term in a female elephant. The elephant would have a wooly mammoth. This clever technique makes moot the prior thinking that a mammoth genome would need to be synthesized in the laboratory. No need to do this (and we can't anyway), because we've got the elephant's cell, and with the mapping of the mammoths DNA, we can translate the one to the other.

So far so good, but there's (or was) a hitch: 400,000 changes are a lot of changes, and the process will likely be arduous to the point of not feasible. Enter the "454 machines", which automate a revolutionary new sequencing technique, that, in effect, let biologists do the genomic modifications in batches. According to George Church, genome technologist at Harvard Medical School, about 50,000 "corrective DNA sequences" can be injected into the cell at one time. In this case, with only a few iterations the machines could inject the entire set of necessary modifications, making the science-fiction like scenario a reality.

The cost estimate for the wooly mammoth project is about $10,000,000, which, while not chump change, is a figure that gaurantees that someone with deep pockets and an interest in our archaelogical past will see things through.

As if this isn't zany enough, there are efforts underway to regenerate the Neanderthals, a hominid race closely related to homo sapiens (us) that lived approximately 200,000 to 45,000 years ago, inhabiting Europe, and possibly coexisting with our direct Cro Magnon descendents. No one knows, conclusively, why the athletic, possibly dim-witted, Neanderthals died out those thousands of years ago. We don't know whether they could talk, or to what extent they created a culture similar to early humans (there is evidence that they drew paintings, suggesting an ability to communicate abstractly). What is certain is that, if the sequencing techniques work on the wooly mammoths, there will be no scientific reason that they can't likewise be applied to generating Neanderthals, if (or when) the extinct species' full genome is recovered.

Work on The Neanderthal Project is well-underway. Svante Paabo of the Max Planck Institute for Evolutionary Anthropology, for instance, has been diligently reconstructing the DNA of Neanderthals using bone fragments discovered in Eastern Europe. With the help of the new "454" sequencing machines, he -- along with a similar project at Lawrence Berkely National Laboratory -- expects to get a complete Neanderthal genetic blueprint. In this case, just as with the use of elephants to birth mammoths, a Neanderthal could be delivered from a human female, or (perhaps less ethically questionable), a Chimpanzee.

As evolutionary biologist Hendrik Poinar notes, “The reality is it will happen,” ... “Twenty to 30 years is the span people are talking about.”

And what then? When a creature so like us -- but so different -- walked again among us, what then? Dartmouth College ethicist Ronald M. Green's comment is as creepy as it is probing:

“This was a species we competed with,” ... “We would not want to recreate a situation of two competing advanced hominid species.”

We may just find out.

Stuck on a Hill

In artificial neural network (ANN) research, there's a well known problem of local minima (or maxima). I've worked a bit with ANNs but much more with a (superior) learning algorithm, Support Vector Machines (SVMs). Unlike the latter, ANNs require heuristics to "converge" on an optimal solution given some very large decision surface. The heuristics, to simplify a bit, are intended to get the algorithm to converge on a global, not local, solution. Local solutions are unwanted because they can appear to be global (good) solutions, given some snapshot of the decision surface, but in fact are very bad solutions when one "zooms out" to see the larger picture. This phenomenon is perhaps best illustrated with geographical imagery. If I am walking up and down hills, en route to a very large mountain, a local maxima might be the top of a foothill. But it would hardly be a global maxima, like the top of the mountain. The point is that ANNs can converge on the foothills, telling us that it's the mountain.

ANNs notoriously suffer from this limitation, but the "blindness problem" is endemic to all statistical learning algorithms, including SVMs (though, at least in theory, you can get an optimal solution with an SVM). Using such algorithms to learn from past experience (i.e., training data), you generate an approximation function for the underlying distribution you're trying to model. You see the function as worth the time it took to gather training data, select the features, and train, if it approximates pretty well the underlying target distribution you're interested in. You can tell if it pretty well approximates the underlying distribution if it keeps getting things right, and you don't have to keep making excuses for it (say, by saying that "it's complicated").

Anyway, we can view ANNs, SVMs, and other statistical learners as essentially inductive systems, in the sense that, given a set of prior examples, they learn a rule (classifier) that allows us to predict something about new, unseen examples. They generalize, in the sense that unseen examples that match a profile (learned from the training data), even if not an exact fit, may still be classified correctly. It's not a simple one-to-one match. Hence, generalize.

The problem with the generalization performance of all such systems is two-fold. One, they're limited by the set of features that were chosen (a person selects features that are relevant, like "is followed with 'inc.'" for classifying organization mentions in free text). Two, even given optimal feature selection, the generalization performance of inductive systems is always hostage to the information in the training data. We collect examples to use to train the system, and, in the end, we hope that the training data was adequate to model the true distribution (the thing we really want to predict). Problem is, whatever hasn't occurred yet in this true distribution, can't possibly be collected from past examples, and so the entire approach is hostage to the fact that things change in real-life environments, and the larger "pattern" of the true distribution as it unfolds in time may not be captured in the training data. Whenever this happens, the approximation function does not model the true distribution. Bummer.

Now, a feature in particular of such inductive systems (we can substitute "supervised statistical learning sytems" if it sounds more fancy) is this local minima or maxima worry, which I introduced with regard to ANNs, but which is really just a handy way of introducing the general problem of generalizing from past cases to future ones writ large. And it is a problem. Consider time sequence prediction (as opposed to, say, sequence classification such as the well-known document classification task in IR research). In time sequence prediction, the goal is take a sequence of elements at time t, and predict the next element at time t+1. Applying this multiple times you can predict a large sequence of elements through some time n.

And this is where the inductive problem comes in, because if the data you're using to predict the next elements came from some set of prior elements, it's possible that the prior elements (your training data), gave you a model that will get you stuck on a foothill, or, will see the top of a small hill as a bottom valley, and so on. You can't be sure, because the future behavior of the true distribution you don't have. And this is why, in the end, induction--however fancy it gets dressed up in mathematical clothing--not only can be wrong, in theory, but often is, in practice.

We can't see into the future, unfortunately. If we could, we could fix the inductive problem by simply adding in the additional information about the true distribution that we're missing in our training data. But in that case, of course, we'd hardly need the approximation.