Before you get started, be sure to check out Part 1 about the basics of control theory.
Engineers tend to have a very realistic view of modeling. Models help provide the fundamentals necessary to design a suitable controller and provide some level of predictive ability. But the model does not do everything for us. Designing a clever controller is almost as much an art as a science. It takes years of experience designing controllers to know what adjustments can be made to improve the dynamics of the controlled system. Models provide a great starting point and are useful for a large portion of the design. However, you can only get so far with the model before you have to make adjustments to the control design, followed by continuous testing and tuning.
Engineering is the application of hard sciences. Since the laws of physics don’t change, it is possible to control physical systems and reliably attain a desired output. In the “soft sciences” where laws are not quite as fixed as they are in physics, researchers are captivated by the challenge of modeling complex social phenomena in order to better understand them. By defining a complex system with a set of mathematical equations, social scientists hope to paint their field as one studying eternal truths, as opposed to just observing the fickle behavior of bipedal primates.
Of course, modeling is a bit of a challenge even for relatively simple systems, so researchers who model complex systems must make sweeping assumptions, even if the object of study is primarily governed by unchanging physical laws, such as the climate. The negligible effects of these assumptions can only be verified by testing on the actual system. A model is only as good as the level of verification. Therefore, models that represent vast complex systems, such as the climate, the economy, or disease spread, are only as reliable as they are verifiable. Since we can’t test these models on the actual system, these types of models are not very reliable at all.
This doesn’t mean that the models are useless. The purpose of the model is to give an idea of how a system likely acts, which has implications for predicting how it behaves under certain circumstances. The mistake that many activist scientists make, one which politicians and pundits regularly use to promote their agenda, is that they assume that the unreliable, unverifiable models can be used to justify some scheme for control. And in incredibly complex systems such as the climate or the economy, there is one variable that is incredibly difficult to control for but which the grand designers love to use as a control variable: human behavior.
This is where the justification for using models to determine public policy falls apart. In engineering, a model is only as good as the level of verification. The validity of a model can only be determined by testing on the actual system, and you are more likely to win the lottery than have your product work on the first try after implementing your control design. Complex systems like the climate or the economy can never be observed in a controlled environment in the same way that a motor can. It is incredibly difficult to reliably change only one input in these systems while holding all other variables relatively constant. When that one input is as fickle as human behavior, the difficulty is compounded even further.
The only way to determine the validity of these sorts of models is to compare their performance against actual data. But this only displays the predictive power of the model, not the stability of the system while tweaking control variables. In order to confirm this kind of robustness, it has to be tested on the actual system. Again, for incredibly complex models, such as those used for climate, the economy, or disease spread, this is impossible.
In 1949, economist William Phillips created the MONIAC machine, a hydraulic computer that could run the calculations of a circular flow model, a macroeconomic model that demonstrates how money moves through society between workers and employers. In Phillips’s contraption, water was used to represent this flow of money and fish tanks represented different sectors of the economy. For instance, the primitive computer could calculate the economic effects of changes in saving habits. Activating the savings lever on the machine would open a valve, draining water from the expenditure stream. As another example, taxation was represented by pumping water back into the “treasury” tank. Increasing the flow would represent increased tax revenue, while decreasing the water flow would represent decreased tax revenue.
Phillips’s work in macroeconomics was useful for conceptualizing certain aspects of the economy on a macro-scale. But it also captured the imaginations of economists and policy makers. Just like pulling the levers of the MONIAC could result in a certain outcome, they envisioned achieving desirable outcomes merely by fiddling with the levers of the economy.
Putting aside the ethical question of what constitutes a desirable economic outcome and who should make that decision, and assuming that the economic models are accurate, there are a few key components missing from these grand designs to control our lives that any good engineer today should be able to identify.
Where is the feedback loop?
And where is the controller?
The answer is of course, that there are none, or at least, not very efficient ones. The feedback loop is the democratic process, which takes place every couple years (imagine if the sensor on your washing machine had a resolution of a couple years). The controller is a hodgepodge of politically appointed bureaucrats and politicians, who happen to be appointed by the democratic feedback loop. I’m not aware of any real world engineering applications where the control parameters are determined by the information coming from the feedback loop. It sounds like a recipe for disaster.
This is one of the myriad of practical problems in thinking that models justify the control of human actions to achieve a desirable outcome. Academics who say that their model shows that setting reference “x” will result in desirable outcome “y” must also be able to account for a feedback loop and a controller. Absent these two things, you will know that the creator of the model has no actual experience in controlling a complex system in the real world. They are merely selling themselves as saviors of mankind, whose theories will lead us to utopia, if only we do exactly what the model says.
In summary, models are a great tool to help conceptualize the workings of complex systems. They also can allow us to predict outcomes when we know the values of certain inputs. Models are a useful starting point for control design but are insufficient on their own. Controllers need to be simulated, tuned, and repeatedly tested and modified to ensure robustness.
In fields that study incredibly complex systems such as the climate, the economy, or the spread of diseases, models alone are not sufficient to determine whether a desired outcome is possible merely by changing the reference value at the input. The models for these massive systems are impossible to test and tune in a real world situation. Not only that, but the models used to justify certain methods of control do not even take into account the basic building blocks of a control design such as a feedback loop and a control algorithm. Suffice it to say, modeling of large, complex systems may be a useful tool for better understanding these systems, but it is an unreliable basis for control to attain certain desirable outcomes.