People particles

John Hawks responds to the new issue of Science with its focus on complex systems, networks, and social science. It seems that, overall, Hawks sees the value in agent-based modeling and complexity. Individual differences are not the only determining factor of behavior and network effects have to be considered.

I do share his criticism though. That often models assume homogeneity among all the agents. There are models that embrace heterogeneity and investigations of the effects of heterogeneous populations on collective outcomes is actually a specific topic for many complexity researchers now. The other criticism, that any modeling approach that treats the units as particles be called 'physics', is incredibly valid. I despise the term 'socio-physics' or 'econo-physics' - I think they're ridiculous names. They are the result of the group of physicists who crossed disciplines and decided to come up with a new name to distinguish themselves from social science. Well, Hawks also makes a remark about that as well (regarding disciplinary boundaries).

Anywho, I would highly advise against calling agent-based models as "particle models". Agent implies something very different than particle does. The term particle, to me, makes the agents in the system seem very passive, while in many models its key that the agents show, well, agency and have intentions, opinions, and tastes.

I'm glad to see complex systems making their way into the mainstream, although I had hoped I would be embedded more fully in them before that happened. Like catching a wave at the right time.

Gallery of 2-d classification

Through the Machine Learning Newsgroup I was pointed to a gallery of 2D classifiers.

There are a number of strange or simple classifcation problems such as this one:



And Fawcett (the author) has trained a variety of different classifcation algorithms to try and detect the pattern. Most social scientists will be familiar with the logistic regression. You can see on the above example that logistic regression will fail horribly:



Then, one of my recent machine learning interests, support vector machines, does a bit better:



In any of linear cases logistic regression performs fantastic. While some of the other algorithms will get messy or incorrect results using a small N, the logistic will get an almost perfect classification. This page is a great resource for exploring the power and limitations of several classification methods.

Special Issue of Science: Complexity

The July 24th Science is devoted to complexity and network science. At every turn in complexity and network science we see physicists, computer scientists, and social scientists in cahoots. Science also has a piece on how network science on its own is growing as its own discipline replete with career paths and funding. Barabási looks friendlier than I thought he would.

New book

I received a book yesterday, finally. I had ordered John Urry's Global Complexity over a month ago, but it was out of print for a short time. I had to wait for them to print me up a fresh copy, I guess. I'm excited since it's exactly the perspective I'm looking for on international interaction.

Often before I read a book I like to read the comments that other people write about it on Amazon or other sites. I usually read the comments from those who rated the book low as they often have the most striking opinions. The lowest comment from this book is:

Urry seemed to condense what should have been 600 pages into 150. And they were some of the hardest 150 pages I've read. Complexity theory is a fascinating topic and its application to globalization is definitely relevant. However, I found Urry very difficult to follow and I was left unconvinced. I would strongly suggest reading M. Mitchell Waldrop's book "Complexity" to get a much clearer perspective on the theory.

He notes that the book is dense as though that were a problem. I prefer dense books. Just give me the idea so I can mull it over. I don't need the author to do all the thinking for me. If Urry had expanded this book to 600 pages I probably wouldn't read it.

I'm a slow reader. Not because my actual reading speed is slow, but because I often stumble on an interesting idea while reading and trail off into thought. I prefer if the reading is dense so that it can accommodate my flights. All else is fluff. With books like Waldrop's* I find myself quickly bored and skimming for the next interesting part. Urry's book seems densely interesting. Looking forward to it.

* - Although that book is actually very interesting - there are better examples for what I'm talking about.

Do you enjoy looking at graphs?

I saw a graph made from age distribution data from the Census Bureau at Economist's View. It's a pretty cool chart. Except for the fact that the default Excel coloring and labellings are uglier than sin.



Excel has some incredible visualization capabilities if you know how to use them. I thought that such interesting presentation of interesting data should not go to waste so I produced my own version also in Excel:



It's like geological strata on the side of the highway.

Free downloadable book on computational modeling

It's mighty scant at this point, but it includes plenty of problems for people who might want to challenge themselves and improve their understanding of the topics. And it's at Version 0.0.10... it has a ways to go. It's all done in Python too, for anyone who might be interested in Python.

I'd probably do everything in R though.

Its the simple things

vs.



I found it really difficult to tell when the trend was negative or positive without a good 0-line. Also, you probably only have to label every few month labels, like Jan., Apr., July., Oct. - the current axis makes everything all crowded.

Via Traffic