Extracting symbolic knowledge from recurrent neural networks—A fuzzy logic approach

Abstract

Considerable research has been devoted to the integration of fuzzy logic (FL) tools with classic artificial intelligence (AI) paradigms. One reason for this is that FL provides powerful mechanisms for handling and processing symbolic information stated using natural language. In this respect, fuzzy rule-based systems are white-boxes, as they process information in a form that is easy to understand, verify and, if necessary, refine.The synergy between artificial neural networks (ANNs), which are notorious for their black-box character, and FL proved to be particularly successful. Such a synergy allows combining the powerful learning-from-examples capability of ANNs with the high-level symbolic information processing of FL systems.In this paper, we present a new approach for extracting symbolic information from recurrent neural networks (RNNs). The approach is based on the mathematical equivalence between a specific fuzzy rulebase and functions composed of sums of sigmoids. We show that this equivalence can be used to provide a comprehensible explanation of the RNN functioning. We demonstrate the applicability of our approach by using it to extract the knowledge embedded within an RNN trained to recognize a formal language.

Rule extraction from autoencoder‐based connectionist computational models

“…As stated before, the rule extraction approaches that can be applied to discover the hidden knowledge from the ANN-structured computational models are limited to those that carry out classification tasks. [19][20][21][22][27][28][29] In addition, other available rule extraction techniques aim at classification problems, 20,21,[30][31][32][33] but there are no reports of their application to nonclassification mapping problems.…”

An expert fuzzy cognitive map for reactive navigation of mobile robots

“…In other words, weights of decision concepts for production P i+1 are obtained from available information at production P i including current weights of factor concepts and causal links. Selection of motion concepts is the most delicate part of the algorithm [30]. There are n = 15 factor concepts which are the input fuzzy states as given in Table 2, and another m = 4 decision concepts for the wheels' velocity controls, i.e., for setting the left wheel velocity to low (LWVL) or to high (LWVH), and for setting the right wheel velocity to low (RWVL) or to high (RWVH).…”

Adaptive fuzzy multivariable controller design based on the Lyapunov scheme

“…Based on this linguistic information and by using computing with words idea, we construct two fuzzy approximated functions for f(x) and g(x) denoted by ) ( x f and ) ( x g , respectively. There are some systematic approaches to do this; one of the possible approaches is the use of neural networks [2].…”

“…In many applications of the fuzzy rule-based systems, fuzzy "if-then" rules are heuristically obtained from human expert knowledge. The use of neural networks to extract the fuzzy rule-base is investigated in many papers [1], [2], but stability of the closed-loop system can not be shown and guaranteed in this kind of methods. The Modelbased fuzzy control approach [3] is another possibility to solve the aforementioned problems.…”