Last years I have been interested in complexity science. Complex systems are systems where the probability for unexpected events to appear is not null.
Control theory is usually based on precise mathematical laws, for instance, classic control theory is based on the assumption of any system as linear. Applying this simplification it is possible to design the classical PID controllers. System controllers are useful preserving the stability of the system, and driving it to a desired working point.
The basement of those control theories is consider that the systems are deterministic. Even if they are not linear, we can find a precise law to predict their behavior. However, real physical systems are usually working under uncertainty. Although in this new case we can follow similar approaches trying to optimize the most probable states, when the complexity of the systems increases (due to a huge number of variables with uncertainty associated) we cannot find any optimal solution.
Engineers, experts in automatics, found AI as a way to improve the control of systems under unexpected events that linear control cannot manage. It is well known from many years the use of AI for supervisory control of industrial facilities. The initial successful attempts were done through computational expert systems. These systems are built through a set of strict cause-effect rules (that follow an IF THEN scheme) obtained from a human expert that know the process to control trying to imitate his reasoning process that it is known as knowledge base, and a software program that manage the processing of that rules that it is usually known as the inference engine.
Expert systems can cope with events unpredictable for the classic control model but expected by a human expert. They work changing the reference signals for certain detectable events following rules stored at the knowledge base. This approach fits better the real physical systems, however, it does not fit well complex systems because it only can cope with expected events that a good expert can know, and on the other hand, the rules of a complex system can be fuzzy. Experts that supervise a system follow imprecise behaviors, they follow rules similar to: “If the temperature is high and it is growing fast, reduce the fuel admission a little”.
Fuzzy logic is integrated in expert systems with inference engines that can process it. This fact lets that these tools can work with rules under a certain degree of uncertainty. This make that they fit better complex systems solving the latter inconvenient, however, the former one, the inability to cope with unexpected events cannot be solved easily with this kind of systems.
Here is where learning seems to be important. A complex system requires supervisory systems that can learn and it can be adaptable to new situation, forgetting fixed and strict rules, as human brain usually works, but the advantage of a computer system over a human supervisor is its capability to provide fast analyses and analytical responses based on measures and data.
I think that in the future, artificial neural networks can be a solution for this kind of supervisory control as a single solution of managing the learning of a modern new expert system paradigm, however, today these technologies are not developed enough to provide the desired results for most of complex systems. Today we are advancing in the comprehension of complex systems and AI algorithms.
Many applications are working every day interacting with the unexpected behavior of multiple human operators and decision makers, in different areas of our society and sectors of the economy, from leisure applications as computer gaming to the most serious investments at Wall Street. Of course, in order to research new approaches of AI controllers working in an environment with multiple human decision makers in competition, I would suggest using as a test field a multiplayer war or race computer game instead of testing the new technologies directly at a computer processing buying and selling orders for the stock exchange, because people that are killed accidentally by a “terminator” at Wall Street cannot be as easily revived as in a computer game, however, Wall Street is being yet a test field for this kind of research work.
Only the well tested applications and paradigms should be deployed in that kind of real scenario.