First meeting of CSS Econ group

This Friday is the first meeting of the CSS department’s econ-modeling group. Interesting group of people and ideas.

Something Rob said in the CSS department’s “new student welcome” has stuck with me. Roughly:

“More and more today we have big data, big information, and big computers. Computational Social Science is all about trying to incorporate these things into a consistent and rigorous model.”

We’re all talking about what we did this summer, and so far, that fits. I’m looking forward to this semester.

Grad Student Pep Talk

I heard a great “pep talk” at the department meeting this last week about “how to succeed as a grad student.” I’ve heard a bit of it before here and there, but I think I’m going to particularly try to achieve a lot of these this semester.

Highlights included:
– “Read a paper a day!”
– “Keep track of your papers with an online bibliography!”
– “Use versioning for everything!”
– “Write everyday!”

My “five papers external-to-project-or-classes” goal this semester is the first above. I think I’ll use Mendeley for the second, along with this blog. I’ve been hearing great things about Git for a while now; I’ll aim to learn that with my class projects for the third. The fourth is more for students in the dissertations phase, I think, but I’ll try to employ it as I get classwork done.

Updates as I go!

Update and a New Semester

I’ve been gone from blogging for a while. My summer was busier than I expected — internship and a project at GMU. Good times!

I now have a backlog of posts I banged out over the summer; I need to get those up.

School starts again tomorrow. I have a goal this semester of reading at least 5 non-class-related papers each week — ideally 7 (one a day!) but we’ll see how the semester goes. Hopefully my new Kindle will make this easier — thanks Mom and Dad! Ideally I’ll post notes on each paper as I read them; this way I can record some thoughts somewhere and hold myself to it.

Papers each week should fill out some or one of these categories:
1. Recommended reading from Rob.
2. Reading of “people whose work I’m interested in.”
3. GMU project reading.
4. “Current Journals/Literature.”
a. Current Journals
b. Current literature “in my field” — ABM-related.
5. My own projects.

1. and 4. roughly correspond to “Classics and current literature,” which I think is the job of every grad student to learn.

Roughly this is:

• Learning the “field right now.”
• “Learning the historical literature.”
• GMU project reading.
• “Keeping up with the current literature and research.”

 

Reading this week:
1. Bullard, Duffy (1998), JEDC : A model of learning and emulation with artificial adaptive agents
2. Howitt, Ozak, (2009), WP: Adaptive Consumption Behavior (18pp)
3. Marimon, McGratten, Sargent (1990), JEDC: Money as a medium of exchange in an economy with artificially intelligent agents (38-78pp — lotta plots)
4. Farmer, Geanakoplos (2009), WP: Hyperbolic discounting is rational: Valuing the far future with uncertain discount rates (18pp)
5. Seppecher (2011, forthcoming), MD: Flexibility of wages and macroeconomic instability in an agent-based computational model with endogenous money.
6. Bonus if time: Nourry, Venditti (2010), MD: Endogenous business cycles in OLG economies with multiple consumption goods
7. Bonus if time: Farmer, Lo (1999), Proc. of Nat. Acc. Science: Frontiers of finance: Evolution and efficient markets.

I’ll update as I read this week.

Collection of Links: Folks Doing Multi-Agent Modeling

Just a collection of places and resources for multi-agent modeling. Very incomplete.

Websites:
http://www.agent-based-models.com
— see the excellent “researchers” and “resources” section. I need to spend a lot more time exploring this.

 

Journals:
Computational and Mathematical Organization Theory
— see the “Call for papers.” Cool.

 

Conferences:

 

Centers:

 

People and Papers:

Scientific Computing, Part 1

Today: “Version control for scientific computing,” or, “how not to waste half your research time remembering what you changed in your code.”

As I’ve started to code perhaps-dissertation-worthy projects, I’ve been thinking about how to avoid wasting extensive time in “code-confusion.” If you’ve programmed any not-one-time-and-small project, you’re familiar with this — “is this chart from the new code? I mean the new-new code, not the old new-code…shoot…”.

It’s been a while since I considered the “software engineering process” more seriously than opening Vim and banging out code. As I move from one-off projects (i.e. HW and class projects) to longer-term scientific computing (i.e. potential research topics), I know that won’t work any more. I need careful version, ideally with some sort of scientific-computing bent, if such a thing exists.

A few Google searches netted some very useful articles, conversations, and blog posts. I don’t have time to do the conversations justice now, but my ridiculously incomplete summary is: (1) USE VERSIONING for your scientific computing. Think about including documentation and notes and thoughts with your code. (2) git looks really nice for a combo of “get running fast, doesn’t interrupt workflow too much,” (3) I still need to figure out where to host… sourceforge, google code, and github are the three main options I’m vaguely thinking about.

Quickly throwing up articles, in order of my favorites:

• “Where’s the Real Bottleneck in Scientific Computing?” from the American Scientist, and Software Carpentry
  — This is tied with the next for my favorite discovery. The article is well worth reading, and a quick search turned up the course.
• “Google Code/Sourceforge/GitHub/SciForge as a Science Repository” blog and discussion
  — Blog post about using google code/etc as project management: docs and data as well as code. This lead me to a very helpful discussion at Friend Feed; both have helpful links. Ideas in action at this post.

• “Version Control in Scientific Computing”
  — Nice “introductory” blog post with descriptions, discussion of CVS vs Git. Convinced me to try Git first.

• “How to Complete your PhD”
  — Good blog post and advise; first item is version control.

• A voice for CVS-style
  — Blog post about CVS-style version control. Only read a little; broadly applicable observations in the intro.

• Another course on good scientific computing practices.

“Where’s the Real Bottleneck in Scientific Computing?” from the American Scientist, and Software Carpentry

This is tied with the next for my favorite discovery. The article is well worth reading, and a quick search turned up the course.

The utility of multi-agent modeling

One must always ask why one is employing a particular tool for a job. Your affinity for the tool is not a good primary reason to use that tool — it must provide something particularly useful in answering a question. This goes doubly for complex tools — one must always justify not using a simpler tool.

One broad usefulness of multi-agent models is the ability to add enormous detail to a particular model. Now, enormous detail is a double-edged sword. In my mind, every bit of detail one adds is a potential systematic error introduced into the workings of the system, and this must be given careful thought.

When I say “enormous detail,” I mean both in terms of data and theory. A lot of micro-data exists, and there is a lot of microeconomic theory that I think is a long way from making it into economic models, macroeconomic and otherwise. A related topic is the incorporating real geography into an economic model — I think this area is ripe for research.

Not that implementing these things well will be straightforward in a multi-agent framework, but I think it is feasible. I will report some examples as soon as I have some fleshed out a little more.

Satisficing and searching for elements of a “core” of ABM

I’ve been thinking about what might constitute a mathematical and/or statistical “core” of ABM. On a related note, something people discuss often when talking about agent-based models is employing non-optimizing preferences for agents — the idea being that people don’t optimize in real life, for a number of reasons. I’ve always been curious about alternative models of preferences — prospect theory, for one, or “satisficing behavior.” I just ran across this earlier today, almost entirely on accident:

http://www.moshe-online.com/satisficing/

Very interesting. The article has a little bit of a harsh tone to it, but it sounds very thorough. I’m looking forward to reading more.

I’ve often wondered if “satisficing”-style preferences could be formulated by simply considering “the correct additional costs” in some way or another.

This is subject for another post, but I understand why many people like optimizing (to use the term incredibly fast and loose!), while at the same time understanding why many people really dislike optimizing. Good points on both sides. Something that will need to be solved, I think, with anything that is non-optimizing, is having a good understanding of how and what non-optimizing means for agent behavior, and why that choice of actions is a good approximation to reality. There is an *enormous* amount of “raw understanding” of the mathematical implications of various maximizing-behavior formulations, which is *very good to have* — not because this correctly describes how the world functions, but because any model is necessarily an abstraction from reality (“is false,” to quote Dr. Box), and we must understand how the false things we have created *act* in their possible existence-spaces. If we don’t understand how our creations act in their created space, how will we know when they are deviating from reality in ways we strongly care about?

In my very limited view, the thing “optimizing” has going for it strongly is the thorough understanding that we have of this agent’s “action space.”

Yes, yes, this is all ridiculously vague. Definitely thoughts for future posts…

Many thoughts, all from podcasts

I’ve been catching up on some of NPR’s Planet Money podcasts as I get my weekend going. I forgot I enjoyed these so much. In one of my favorite papers of all time, Hal Varian (who wrote one of the foundational <a href="http://www.amazon.com/Microeconomic-Analysis-Third-Hal-Varian/dp/0393957357"micro textbooks that I’m told the “previous generation” of economists worked through) wrote that “you should look for your [research] ideas outside the academic journals—in newspapers, in magazines, in conversations, and in TV and radio programs.” These planet money podcasts reminded me why.

No time now for exposition on that, but I do want to quickly list the podcasts with “seeds of ideas:”

• Discipline vs forgiveness in borrowing, and what that means for growth.
  • I haven’t finished listening to this yet. One quick stab — perhaps the “weight of results” of making very complex calculations should be laid on those most able to make those calculations. Super-quick version: they seem to say, “Spain lays the consequences of mis-understanding a complex calculation of a person/business’ potential insolvency primarily on the person, vs in the US, where a lot of that is laid on the lender.” Quick-reaction is: well, in any transaction, both bear responsibility. With these laws, either version, that responsibility is necessarily laid more on one party vs the other; policy-makers must simply choose which party bears the greater weight. We want to encourage the best calculations the be made, so perhaps they should lay that responsibility on those most able to actually make the calculation — not because it’s fair or not fair or whatever, but because that makes the system self-balancing and better-working. Thus in the US, the lenders, who ostensibly have it as their job to know creditworthiness of a business/person/etc, — because that’s their job, that’s what they do — perhaps they should bear the weight of the complex calculation of creditworthiness. And here “bearing the weight” means they are left holding the bag if it doesn’t work — they bear the risk. “And thus that system encourages more growth,” since you have people specialized in allocating resources actually allocating the resources. People may say, “well obviously that didn’t work in the US with Financial Crisis,” but that ignores that I think (big “I think” here) that Spain actually has it much, much worse, debt- and production-wise, and has for many years. On an unrelated note, when I started this podcast, I thought of the Rocky quote, “…I figured if I break the guy’s thumb, he gets laid off, right? Then he can’t make payments…” …never seemed like a good bookie strategy.
• Inside the Great Depression
  • Great story. I was struck by the observation, halfway through, that fruit was going to rot on trees while kids in cities starved, because no one had money to spend. I listened to this right after the “Gold Standard, Part II” podcast, and it fit so well. Interesting that the fellow making that observation in this podcast was still strongly in favor of the gold standard, even as he recognized how it was affecting the Depression almost exactly how the “Gold Standard, Part II” folks talked about it happening.
• Production vs rentier states, regarding Libya.
  • They talk, at the end about the difference between Haiti (not) recovering after revolution a while back, vs the potential for Libya to recover if Gaddafi is outed. One asks — what was happening in Haiti, after that governmental overturn a while back? One wants to dive into the details of what makes a modern city of any particular size run. One wants to try to trace out the details of what makes/goes into any particular persons job funcitoning. Ask this in some quantitative way — how many people go into producing the goods I use in my daily life? What is the “production chain” for an average good or service in my city? …certainly a lot of this is unknowable with current data. Network structures would figure into this, perhaps. More thoughts later.
  • Q2 is, could one model Norway’s approach?
• The gold standard, Part 1 and Part 2 (there are more but I haven’t heard the rest)
  • Many thought-provoking conversations here. I’m tempted to go pick up my copy of The Great Contraction again. The thing that catches my attention here is that these arguments discussed are all essentially verbal models. One wonders what minimal requirements on agents and environments would be needed to support the verbal models laid out here. Part of me suspects that there are inevitably “things left unconnected” in the verbal models — that they aren’t a closed system. This, of course, is one of the reasons formal mathematical and computational models are built: thinking through the math or the computation forces you to face the “loose ends” of your system. (Of course the models must be built well…)

    This also reminds me that I think there may still be a lot of work to be done in building and exploring any sort of non-strictly-optimizing agents, as I suspect the fellow advocating the return to the gold standard might prefer. This is a future post.

I need to figure out how to write less, I think. This post didn’t turn out to be “quick thoughts,” whoops. And I still need a haircut today!

Posts I would like to write

Hello all. I’ve been trying to decide how to handle the fact that I keep wanting to “link forward” to “posts I will someday write.” I have posts in my head, but not enough time to write them — what to do?

My quick solution is to just list them, here, and link to this page — and — link back to the page where I reference my non-existent posts 🙂 Thus hopefully not forgetting, in the future, where those links are. We will see how this work.

Future links:

• “Being realistic” and “understanding your model”
  • I don’t think there is a tradeoff here — just more work. In modeling we face a dual problem: we are trying to understand true behavior of actual people, and we must also understand the artificial behavior of our artificial representations of people. Any agent is necessarily artificial. When we use them we must understand how they act in their own right — at least with respect to whatever activity we care about. (If I am modeling home purchases, maybe I don’t care if the person likes action movies or romantic comedies.)
  1

Rational Expectations

I’ve found myself in multiple conversations lately about rational expectations. I decided I would add Rational Expectations to my long and growing reading list. I was given this by someone a long time ago, but never read it.

The first chapter already has very interesting tidbits. It mentions rational expectations as a “controversial set of ideas that many economists view as having little policy relevance” (to paraphrase somewhat). Interesting. Perhaps the first edition shows its age there a little? Also, I never knew that Muth developed his idea of rational expectations while collaborating with Herbert Simon, the “father of bounded rationality.” Apparently they were working on the same research model and both wrestling with how to model expectations, which could be a significantly important influencer of system dynamics. There’s an interesting quote by Simon related to the usefulness of something along the lines of rational expectations.

I’ve been going back and forth in my mind about what role rational expectations (RE) might play in agent modeling. I heard a great talk the other day, and the speaker noted that “you always need to remember that the simplest human is smarter than your most intelligent agent.” The best justification I’ve ever encountered for RE addresses this exact point: we can never hope to make an agent as broadly intelligent or clever or communicative as a real human, so instead of accidentally making an economic agent too dumb, and ending up with “model artifacts” from that, RE aims to make the agent as smart as he/she possibly can be, and see what results come from that. I think this is an incredibly good point, one that must be taken very seriously.

Of course, one might argue that some artifacts one might get from making an agent “dumber” than RE might be how the world actually is — that some people are less-then-RE, in some important and significant way. But as someone (whose name slips the mind at the moment) once said, one can get lost in the “wilderness of bounded rationality.” There’s only one way to have rational expectations, but infinite ways to be boundedly rational.

I have is two reactions to this, one immediate, and one that’s been emerging slowly. The immediate reaction is that, whenever a type of bounded rationality is introduced in a particular model, it could be very useful to start a model with full measure of “RE agents*,” and explore how the model dynamics change as the percentage of boundedly rational agents is slowly increased.
(* I’ll note that as soon as non-RE agents are introduced, it may become exceedingly difficult to have truely RE agents, and we may need to substitute something approximating that.)

The second reaction is really almost an offshoot of the footnote to the last paragraph. Perhaps a boundedly rational agent should really just be a small step away from a “truely RE” agent. This could be based on any number of things — largely, I imagine, the structure of the game or trading environment. I’m something of a fan, however, of still using “true”/objective probabilities, at least as a leaping-off point. Thus we would retain the “expectations” part of rational expectations. Perhaps this should be called “bounded rational expectations.”

I need to write an example of this. Perhaps that will be the subject of a future entry.