There Are Models And There Are Models
I’m 74, and I’ve been programming computers nearly as long as anyone alive.
When I was 15, I’d been reading about computers in pulp science fiction magazines like Amazing Stories, Analog, and Galaxy for a while. I wanted one so badly. Why? I figured it could do my homework for me. Hey, I was 15, wad’ja expect?
I was always into math, it came easy to me. In 1963, the summer after my junior year in high school, nearly sixty years ago now, I was one of the kids selected from all over the US to participate in the National Science Foundation summer school in mathematics. It was held up in Corvallis, Oregon, at Oregon State University.
It was a wonderful time. I got to study math with a bunch of kids my age who were as excited as I was about math. Bizarrely, one of the other students turned out to be a second cousin of mine I’d never even heard of. Seems math runs in the family. My older brother is a genius mathematician, inventor of the first civilian version of the GPS. What a curious world.
The best news about the summer school was, in addition to the math classes, marvel of marvels, they taught us about computers … and they had a real live one that we could write programs for!
They started out by having us build design and build logic circuits using wires, relays, the real-world stuff. They were for things like AND gates, OR gates, and flip-flops. Great fun!
Then they introduced us to Algol. Algol is a long-dead computer language, designed in 1958, but it was a standard for a long time. It was very similar to but an improvement on Fortran in that it used less memory.
Once we had learned something about Algol, they took us to see the computer. It was huge old CDC 3300, standing about as high as a person’s chest, taking up a good chunk of a small room. The back of it looked like this.
It had a memory composed of small ring-shaped magnets with wires running through them, like the photo below. The computer energized a combination of the wires to “flip” the magnetic state of each of the small rings. This allowed each small ring to represent a binary 1 or a 0.
How much memory did it have? A whacking great 768 kilobytes. Not gigabytes. Not megabytes. Kilobytes. Thats one ten-thousandth of the memory of the ten-year-old Mac I’m writing this on.
It was programmed using Hollerith punch cards. They didn’t let us anywhere near the actual computer, of course. We sat at the card punch machines and typed in our program. Here’s a punch card, 7 3/8 inches wide by 3 1/4 inches high by 0.007 inches thick. (187 x 83 x.018 mm).
The program would end up as a stack of cards with holes punched in them, usually 25-50 cards or so. I’d give my stack to the instructors, and a couple of days later I’d get a note saying “Problem on card 11”. So I’d rewrite card 11, resubmit them, and get a note saying “Problem on card 19” … debugging a program written on punch cards was a slooow process, I can assure you
And I loved it. It was amazing. My first program was the “Sieve of Eratosthenes“, and I was over the moon when it finally compiled and ran. I was well and truly hooked, and I never looked back.
The rest of that summer I worked as a bicycle messenger in San Francisco, riding a one-speed bike up and down the hills delivering blueprints. I gave all the money I made to our mom to help support the family. But I couldn’t get the computer out of my mind.
Ten years later, after graduating from high school and then dropping out of college after one year, I went back to college specifically so I could study computers. I enrolled in Laney College in Oakland. It was a great school. The Laney College Computer Department had a Datapoint 2200 computer, the first desktop computer.
It had only 8 kilobytes of memory … but the advantage was that you could program it directly. The disadvantage was that only one student could work on it at any time. However, the computer teacher saw my love of the machine, so he gave me a key to the computer room so I could come in before or after hours and program to my heart’s content. I spent every spare hour there. It used a language called Databus, my second computer language.
The first program I wrote for this computer? You’ll laugh. It was a test to see if there was “precognition”. You know, seeing the future. My first version, I punched a key from 0 to 9. Then the computer picked a random number, and recorded if I was right or not.
Finding I didn’t have precognition, I re-wrote the program. In version 2, the computer picked the number before, rather than after, I made the selection. No precognition needed. Guess what?
No better than random chance. And sadly, that one-semester course was all that Laney College offered. That’s the extent of my formal computer education. The rest I taught myself, year after year, language after language, concept after concept, program after program.
Ten years after that, I bought the first computer I ever owned — the Radio Shack TRS-80, AKA the “Trash Eighty”. It was the first notebook-style computer. I took that sucker all over the world. I wrote endless programs on it, including marine celestial navigation programs that I used to navigate by the stars between islands the South Pacific. It was also my first introduction to Basic, my third computer language.
And by then IBM had released the IBM PC, the first personal computer. When I returned to the US I bought one. I learned my fourth computer language, MS-DOS. I wrote all kinds of programs for it. But then a couple years later Apple came out with the Macintosh. I bought one of those as well, because of the mouse and the art and music programs. I figured I’d use the Mac for creating my art and my music and such, and the PC for serious work.
But after a year or so, I found I was using nothing but the Mac, and there was a quarter-inch of dust on my IBM PC. So I traded the PC for a piano, the very piano here in our house that I played last night for my 19-month-old granddaughter, and I never looked back at the IBM side of computing.
I taught myself C and C++ when I needed speed to run blackjack simulations … see, I’d learned to play professional blackjack along the way, counting cards. And when my player friends told me how much it cost for them to test their new betting and counting systems, I wrote a blackjack simulation program to test the new ideas. You need to run about a hundred thousand hands for a solid result. That took several days in Basic, but in C, I’d start the run at night, and when I got up the next morning, the run would be done. I charged $100 per test, and I thought “This is what I wanted a computer for … to make me a hundred bucks a night while I’m asleep.”
Since then, I’ve never been without a computer. I’ve written literally thousands and thousands of programs. On my current computer, a ten-year-old Macbook Pro, a quick check shows that there are well over 2,000 programs I’ve written. I’ve written programs in Algol, Datacom, 68000 Machine Language, Basic, C/C++, Hypertalk, Mathematica (3 languages), Vectorscript, Pascal, VBA, Stella computer modeling language, and these days, R.
I had the immense good fortune to be directed to R by Steve McIntyre of ClimateAudit. It’s the best language I’ve ever used—free, cross-platform, fast, with a killer user interface and free “packages” to do just about anything you can name. If you do any serious programming, I can’t recommend it enough.
I bring all of this up to let you know that I’m far, far from being a novice, a beginner, or even a journeyman programmer. I was working with “computer based evolution” to try to analyze the stock market before most folks even heard of it. I’m a master of the art, able to do things like write “hooks” into Excel that let Excel transparently call a separate program in C for its wicked-fast speed, and then return the answer to a cell in Excel …
Now, folks who’ve read my work know that I am far from enamored of computer climate models. I’ve been asked “What do you have against computer models?” and “How can you not trust models, we use them for everything?”
Well, based on a lifetime’s experience in the field, I can assure you of a few things about computer climate models and computer models in general. Here’s the short course.
• A computer model is nothing more than a physical realization of the beliefs, understandings, wrong ideas, and misunderstandings of whoever wrote the model. Therefore, the results it produces are going to support, bear out, and instantiate the programmer’s beliefs, understandings, wrong ideas, and misunderstandings. All that the computer does is make those under- and misunder-standings look official and reasonable. Oh, and make mistakes really, really fast. Been there, done that.
• Computer climate models are members of a particular class of models called “Iterative” computer models. In this class of models, the output of one timestep is fed back into the computer as the input of the next timestep. Members of his class of models are notoriously cranky, unstable, and prone to internal oscillations and generally falling off the perch. They usually need to be artificially “fenced in” in some sense to keep them from spiraling out of control.
• As anyone who has ever tried to model say the stock market can tell you, a model which can reproduce the past absolutely flawlessly may, and in fact very likely will, give totally incorrect predictions of the future. Been there, done that too. As the brokerage advertisements in the US are required to say, “Past performance is no guarantee of future success”.
• This means that the fact that a climate model can hindcast the past climate perfectly does NOT mean that it is an accurate representation of reality. And in particular, it does NOT mean it can accurately predict the future.
• Chaotic systems like weather and climate are notoriously difficult to model, even in the short term. That’s why projections of a cyclone’s future path over say the next 48 hours are in the shape of a cone and not a straight line.
• There is an entire branch of computer science called “V&V”, which stands for validation and verification. It’s how you can be assured that your software is up to the task it was designed for. Here’s a description from the web
What is software verification and validation (V&V)?
820.3(a) Verification means confirmation by examination and provision of objective evidence that specified requirements have been fulfilled.
“Documented procedures, performed in the user environment, for obtaining, recording, and interpreting the results required to establish that predetermined specifications have been met” (AAMI).
820.3(z) Validation means confirmation by examination and provision of objective evidence that the particular requirements for a specific intended use can be consistently fulfilled.
Process Validation means establishing by objective evidence that a process consistently produces a result or product meeting its predetermined specifications.
Design Validation means establishing by objective evidence that device specifications conform with user needs and intended use(s).
“Documented procedure for obtaining, recording, and interpreting the results required to establish that a process will consistently yield product complying with predetermined specifications” (AAMI).
Further V&V information here.
• Your average elevator control software has been subjected to more V&V than the computer climate models. And unless a computer model’s software has been subjected to extensive and rigorous V&V. the fact that the model says that something happens in modelworld is NOT evidence that it actually happens in the real world … and even then, as they say, “Excrement occurs”. We lost a Mars probe because someone didn’t convert a single number to metric from Imperial measurements … and you can bet that NASA subjects their programs to extensive and rigorous V&V.
• Computer modelers, myself included at times, are all subject to a nearly irresistible desire to mistake Modelworld for the real world. They say things like “We’ve determined that climate phenomenon X is caused by forcing Y”. But a true statement would be “We’ve determined that in our model, the modeled climate phenomenon X is caused by our modeled forcing Y”. Unfortunately, the modelers are not the only ones fooled in this process.
• The more tunable parameters a model has, the less likely it is to accurately represent reality. Climate models have dozens of tunable parameters. Here are 25 of them, there are plenty more.
• The climate is arguably the most complex system that humans have tried to model. It has no less than six major subsystems—the ocean, atmosphere, lithosphere, cryosphere, biosphere, and electrosphere. None of these subsystems is well understood on its own, and we have only spotty, gap-filled rough measurements of each of them. Each of them has its own internal cycles, mechanisms, phenomena, resonances, and feedbacks. Each one of the subsystems interacts with every one of the others. There are important phenomena occurring at all time scales from nanoseconds to millions of years, and at all spatial scales from nanometers to planet-wide. Finally, there are both internal and external forcings of unknown extent and effect. For example, how does the solar wind affect the biosphere? Not only that, but we’ve only been at the project for a couple of decades. Our models are … well … to be generous I’d call them Tinkertoy representations of real-world complexity.
• Many runs of climate models end up on the cutting room floor because they don’t agree with the aforesaid programmer’s beliefs, understandings, wrong ideas, and misunderstandings. They will only show us the results of the model runs that they agree with, not the results from the runs where the model went off the rails. Here are two thousand runs from 414 versions of a model running first a control and then a doubled-CO2 simulation. You can see that many of the results go way out of bounds.
As a result of all of these considerations, anyone who thinks that the climate models can “prove” or “establish” or “verify” something that happened five hundred years ago or a hundred years from now is living in a fool’s paradise. These models are in no way up to that task. They may offer us insights, or make us consider new ideas, but they can only “prove” things about what happens in modelworld, not the real world.
Be clear that having written dozens of models myself, I’m not against models. I’ve written and used them my whole life. However, there are models, and then there are models. Some models have been tested and subjected to extensive V&V and their output has been compared to the real world and found to be very accurate. So we use them to navigate interplanetary probes and design new aircraft wings and the like.
Climate models, sadly, are not in that class of models. Heck, if they were, we’d only need one of them, instead of the dozens that exist today and that all give us different answers … leading to the ultimate in modeler hubris, the idea that averaging those dozens of models will get rid of the “noise” and leave only solid results behind.
Finally, as a lifelong computer programmer, I couldn’t disagree more with the claim that “All models are wrong but some are useful.” Consider the CFD models that the Boeing engineers use to design wings on jumbo jets or the models that run our elevators. Are you going to tell me with a straight face that those models are wrong? If you truly believed that, you’d never fly or get on an elevator again. Sure, they’re not exact reproductions of reality, that’s what “model” means … but they are right enough to be depended on in life-and-death situations.
Now, let me be clear on this question. While models that are right are absolutely useful, it certainly is also possible for a model that is wrong to be useful.
But for a model that is wrong to be useful, we absolutely need to understand WHY it is wrong. Once we know where it went wrong we can fix the mistake. But with the complex iterative climate models with dozens of parameters required, where the output of one cycle is used as the input to the next cycle, and where a hundred-year run with a half-hour timestep involves 1.75 million steps, determining where a climate model went off the track is nearly impossible. Was it an error in the parameter that specifies the ice temperature at 10,000 feet elevation? Was it an error in the parameter that limits the formation of melt ponds on sea ice to only certain months? There’s no way to tell, so there’s no way to learn from our mistakes.
Next, all of these models are “tuned” to represent the past slow warming trend. And generally, they do it well … because the various parameters have been adjusted and the model changed over time until they do so. So it’s not a surprise that they can do well at that job … at least on the parts of the past that they’ve been tuned to reproduce.
But then, the modelers will pull out the modeled “anthropogenic forcings” like CO2, and proudly proclaim that since the model no longer can reproduce the past gradual warming, that demostrates that the anthropogenic forcings are the cause of the warming … I assume you can see the problem with that claim.
In addition, the gridsize of the computer models are far larger than important climate phenomena like thunderstorms, dust devils, and tornados. If the climate model is wrong, is it because it doesn’t contain those phenomena? I say yes … computer climate modelers say nothing.
Heck, we don’t even know if the Navier-Stokes fluid dynamics equations as they are used in the climate models converge to the right answer, and near as I can tell, there’s no way to determine that.
To close the circle, let me return to where I started—a computer model is nothing more than my ideas made solid. That’s it. That’s all.
So if I think CO2 is the secret control knob for the global temperature, the output of any model I create will reflect and verify that assumption.
But if I think (as I do) that the temperature is kept within narrow bounds by emergent phenomena, then the output of my new model will reflect and verify that assumption.
Now, would the outputs of either of those very different models “evidence” be about the real world?
Not on this planet.
And that is the short list of things that are wrong with computer models … there’s more, but as Pierre said, “the margins of this page are too small to contain them” …
My very best to everyone, stay safe in these curious times,
This article originally appeared on the science site wattsupwiththat.com]]>