OVER THE last century, what started out as a collection of simple home thermostats and flywheels on steam engines--the science of automation and control--evolved into a powerful tool of progress. We have also passed the age of single loop PID control. We’ve realized that what needs to be controlled is not a pressure or a temperature, but the productivity and safety of multivariable unit operations.
While achieving all this, we have failed to inform society as a whole that this know-how exists and is available to control non-industrial processes. In fact, when people ask, “What do you do?” and I answer, “Process control,” I just get a blank stare. People don’t even know that our profession exists. It was for this reason that my course at Yale was offered by the Chemical Engineering department., and the reason that my publisher is promoting my books as part of its Electrical Engineering series. They don’t even know that we exist!
Yet we not only exist, but we have a lot to offer. For example, the American economy is also a process. It is a multivariable process that is being controlled in a “manual and on-off” manner. The only manipulated variable that is throttled is the interest rate on federal funds, which lowers or raises the flow of money into our economic process. The operator who adjusts the manual valve is named Alan Greenspan.
It seems to me that we could and should do better!
Understanding the Process We Control
Obviously, the processes of both the national and the global economies are multivariable and nonlinear, with variable gain and variable dead times. In industry, we have learned that to control such processes, self-adaptive, model-based ANN (artificial neural network), multi-variable controls are needed. Self-teaching ANN algorithms are trained on past history (the data of the last decades of operation of the process). Such self-teaching algorithms learn the “personality” of the controlled process in a similar manner as a tennis player, who learns the correlation between his or her actions and results, without necessarily understanding the theories of Newton or the laws of aerodynamics.
So, let us look at the multivariable process of the economy, any economy. As shown in the chart below, the controlled variables (the set points) of this process are the gross domestic product (GDP) and the leading economic indicators (LEIs) of the economy.
CONTROLLING THE ECONOMY USING
AN ARTIFICIAL NEURAL NETWORK (ANN)
The ANN is capable of learning the relationships between inputs (manipulated and disturbance variables -- MV, DV) and outputs (controlled variable -- CV) of the economic process based on its past history. Once the model of the process is constructed and "trained," it can be continuously updated by minimizing the difference (em) between its output and that of the economy.
CAPITALIST AND socialist societies might distribute the resulting production differently among the population, but the manipulated and disturbance variables in all economic processes are basically the same. The only difference is that in free societies, most of the variables are allowed to float freely (are in automatic), while in totalitarian societies some of the variables are arbitrarily constrained (are in manual).
The goals of both economies (the set points in the chart above) are to maintain their LEIs and their GDP (U.S. GDP is about $12 trillion/yr) at some desirable value or to keep raising these targets. LEIs are indicators of the standing of the stock and housing markets, percent unemployment, percent inflation (2-3% in the U.S.), increase in wages (-0.3% drop in 2004 in the U.S.), quality of health, education, social security services, etc. In addition, LEIs also consider the value of the currency and the price to earning ratio (P/E) of securities, which during the last 10 years was 26 in the U.S., the worst since 1927. Some LEIs also include the housing “bubble” (market value of real estate in units of years of rental income), which is 17 today in the U.S.
As shown in the chart, the relationship between manipulated variables (such as interest rate, taxation, trade or energy policy) and controlled variables (GDP or LEIs) are functions of the “gains” at the nodes in the “hidden” layer (or layers) of this self-teaching ANN model of the economic process. The ANN controller looks at the difference (ef) between the desired (SP) and actual (CV) values of the GDP (or LEIs) and adjusts the manipulated variable (MV) when an error exists. The ANN controller uses the inverse of the model of the economy to continuously update itself. It does that by comparing its output with that of the actual economy and, if a difference (error em) exists, correcting the model. Naturally, because both the MV and the disturbance variables (DV) influence the real process, they are also inputs to the ANN model.
Before one can use an ANN model of a process as a feed forward predictor, it must be trained on past data of process performance. In case of the economy, the model can be trained on the data of the past decades, just as it would be trained on the historical performance data of a distillation column. From the model’s viewpoint it makes little difference if the energy source to a process is the steam supply to a reboiler or the money supply of the economy.
In both processes, there is a “gain” relationship between the input and the output, the change in the steam (or money) flow and the resulting increase in production of distillate (or GDP). Naturally, the response to the flow of steam (or money) is not instantaneous, but is determined by the time constants and dead times of the processes. In addition, all measurement signals contain some noise. The filter in the chart serves to remove noise. In case of the process of economy, the filter might serve to remove the effects of the arbitrary acts of fund managers, politics, etc.
In the second part of this article, I will describe an ANN model for the economy.
|About the Author|