4
$\begingroup$

On Google Scholar the term fuzzy logic results into a list of 1 million papers, which is a lot. The subproblem of “Fuzzy control” has around 238000. If we are assuming, that the total number of papers in Academia is only 50 million, than the term fuzzy logic occupies 2% of the published information. But how important is the topic for Artificial Intelligence and Robotics? As far as i know, there is a hype around deeplearning. The topic is pushed forward and many lectures are held at universities about machine learning, neural network and stochastic automaton. In contrast, the topic fuzzy control is not covered very well. It sounds a bit esoteric. We have on the one hand a huge number of published papers about the subject, while on the other hand nobody is using fuzzy control in reality. Or at least, it seems so. A closer look into the papers shows, that fuzzy logic is everywhere: from simple line follower robots, over starcraft playing AI engines up to fuzzy nanoparticles which are cruising in the human body.

My question is: is fuzzy logic something which can be ignored because other topics like neural networks are more important or does it make sense to focus on fuzzy control and support the idea of describing a system with linguistic variables?

A similar question was asked by Why did fuzzy logic fall out of fashion? two months ago. But it doesn't contains numbers about published articles.

$\endgroup$
1
$\begingroup$

System architectures containing interconnected semantic networks, convolution kernels, Markov processes, multilayer perceptrons, reinforcement learners, fuzzy logic containers, rules engines, and other AI elements exist and will continue to be developed. It is unwise to consider any one of these ideas completely supplanted by another. Each has its purpose. One may actually be used adjacent to or within another. One may be leveraged to produce the behavior of another, when the other was previously implemented differently.

The reason for fuzzy logic is not difficult to comprehend.

In math proofs, either a step in the proof is valid according to the rules of the mathematical system or it is not. However, even in math, whether a valid step will lead to a complete proof of theorem proposed is unsure. The mathematician may make the step with the expectation that it is the right one and may boldly assert that it is. But such an assertion will occasionally be incorrect. For each step in the proof, confidence and doubt about the value of the choice coexist, even though the mathematical validity of each choice is sure.

Once the steps that prove the theorem are found and confirmed, the theorem may lead to an approach, algorithm, or circuit used in AI. In field use, the crisp bounds of a mathematical environment is replaced with a less precisely known environment. Because of this, in engineering, assumptions are made to simplify the problem space. Even after simplification, any AI system or subsystem in a robot or data center must address the probabilistic nature of observation. The state of external things is only partly known.

What a robotic controller receives as input is a stream of sample vectors, not anything near a complete and accurate representation of the trajectories of every physical object or particle of interest. For instance, in vision, objects may be partially obscured or fully obscured for a period of time. Whatever model is maintained by signals sensing the environment is necessarily an approximation.

Data center analysis processes are similar. The model may instead be related to sales, marketing, the flow of products and services, or something else. The input that maintains the model arises from network messages and transactions, and the result is still an approximation maintained through a series of partial observations.

When the rules of the domain are in doubt, they can be thought of as fuzzy.

Note that this is unlike doubt about which rule to apply as in searching for a mathematical proof, where the cognitive skill of the mathematician is free to gather all information available about the environment of the proof. This brings us to the reason why fuzzy logic should continue to hold its place in AI research and development.

Cognition is fuzzy logic. Cognitive processes can design and document a mathematical proof, but cognition is not itself based on a definitive set of rules. This is why most decisions in business, private life, and government lack a corresponding proof. Logical approaches are weighted in relation to one another and sometimes combined. Pure predicate logic is too definite to represent rules in a complex domain and model-free learning is not yet able to converge in such a way to approximate human cognition.

Is fuzzy logic something which can be ignored because other topics like neural networks are more important ...?

The current interest in artificial networks is not a good reason to dismiss the value of fuzzy logic. One is not a plug-in replacement for the other at this time. Even if someday doubt and logic can be modeled using deep networks, it may be resource inefficient to replace optimized fuzzy logic containers with artificial networks. That determination will have to be made at that time.

The types of components developed through AI research should be considered the building blocks for high level AI designs, and that has been the case. To diminish the variety of building blocks would be to diminish AI. If skewing has occurred in the rate of scholarship across these component types, then AI has already been diminished to the degree that any one component type has been dismissed. This kind of question may help to maintain balance in the parallel forks of AI progress.

Imbalance may be why chat-bots lack cognitive skills. The building blocks incorporated into many production chat-bot systems are not capable of conversation requiring any depth of comprehension. Chat-bots without fuzzy logic may eventually pass Turing's imitation game within a limited context, but business communications has suffered from the replacement of minds that possess cognitive abilities with machines that do not. The result is lower quality customer service.

The remedy is to re-balance research and recognize that cognition goes beyond simple learning and that it is both logical and fuzzy.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.