#### DMCA

## Symbolic Interpretation of Artificial Neural Networks (1996)

### Cached

### Download Links

- [ftp.lans.ece.utexas.edu]
- [ftp.lans.ece.utexas.edu]
- [www.ki.inf.tu-dresden.de]
- [www.ideal.ece.utexas.edu]
- DBLP

### Other Repositories/Bibliography

Venue: | IEEE Transactions on Knowledge and Data Engineering |

Citations: | 47 - 1 self |

### Citations

6465 |
C4.5: Programs for Machine Learning
- QUINLAN
- 1993
(Show Context)
Citation Context ...ead to substantial overheads. ffl C4.5rules: C4.5rules was used by the authors of NeuroRule to extract rules from the iris and breast-cancer databases for comparison reasons. Like ID3 [36], C4.5rules =-=[20]-=- generates decision tree rules based on the available input samples. Therefore, the complexity is moderate, but the performance of the rules generated by C4.5rules is highly affected by the noise leve... |

4282 | Induction of decision trees
- Quinlan
- 1986
(Show Context)
Citation Context ...ring processes lead to substantial overheads. ffl C4.5rules: C4.5rules was used by the authors of NeuroRule to extract rules from the iris and breast-cancer databases for comparison reasons. Like ID3 =-=[36]-=-, C4.5rules [20] generates decision tree rules based on the available input samples. Therefore, the complexity is moderate, but the performance of the rules generated by C4.5rules is highly affected b... |

1061 |
C4.5: Programs for
- Quinlan
- 1993
(Show Context)
Citation Context ...leextraction process becomes easy. œ C4.5 rules: C4.5 rules was used by the authors of NeuroRule to extract rules from the iris and breastcancer databases for comparison reasons. Like ID3, C4.5 rules =-=[28]-=- generate decision tree rules based on the available input samples. Therefore, the complexity is moderate, but the performance of the rules generated by C4.5 rules is highly affected by the noise leve... |

969 | Adaptive mixtures of local experts
- Jacobs, Jordan, et al.
- 1991
(Show Context)
Citation Context ...EX is that it controls the search space through its network while other approaches use heuristic measures to do the same. RuleNet, on the other hand, uses the idea of adaptive mixture of local expert =-=[19] to train -=-a localized ANN then extracts binary rules in a LRE approach. Both RULEX and RuleNet can be classified as "localized LRE" techniques. 2.2.2 Black-box Rule Extraction Techniques Another class... |

841 |
UCI repository of machine learning databases
- Murphy, Aha
- 1994
(Show Context)
Citation Context ...ules relating four binary inputs and four binary outputs. 2. Iris database, a simple classification problem which contains 50 examples each of classes Iris Setosa, Iris Versicolor, and Iris Virginica =-=[32]-=-. These 150 instances were divides into two subsets, the first subset, used for training, is of size 89 and the second is of size 61 and used for testing. Each input pattern has four continuous input ... |

805 |
Mathematica: a system for doing mathematics by computer
- Wolfram
- 1991
(Show Context)
Citation Context ... i;l are the lower and upper boundary values of interval l of input X i respectively. Different discretization approaches can be exploited to compute discretization boundaries of input features X i s =-=[46, 5, 23, 26, 56]-=-. Full-RE uses the Chi2 [25] algorithm 2 , a powerful discretization tool, to compute discretization boundaries of input features. When Full-RE finds more 2 We are thankful to Liu and Setiono for maki... |

432 |
Logic Minimization Algorithms for VLSI Synthesis
- Brayton, Hachtel, et al.
- 1984
(Show Context)
Citation Context ...fication method can be used to perform step 3 of the BIO-RE algorithm (e.g. Karnough map [22], algebraic manipulation, or a tabulation method [29]). We used Espresso 1 to generate the extracted rules =-=[4]-=-. Rules extracted by BIO-RE are represented in the format: If [Not] Input-Variable [And [Not] Input-Variable] \Gamma!Consequent j where: [\Delta] is an optional term and [\Delta] means that the term [... |

304 |
A survey and critique of techniques for extracting rules from trained artificial neural networks
- ANDREWS, DIEDERICH, et al.
- 1995
(Show Context)
Citation Context ... feedforward (specifically, MLP) ANN architectures. A very rich source of 2 RULE EXTRACTION 6 literature review of different rule extraction approaches is a technical report written by Andrews et al. =-=[1]-=-. 2.2.1 Link Rule Extraction Techniques The methodology behind most of the techniques for rule extraction from MLPs can be summarized in two main steps: (i) For each hidden or output node in the netwo... |

225 | Extracting refined rules from knowledgebased neural networks
- Towell, Shavlik
- 1993
(Show Context)
Citation Context ...current node. (ii) For each of these combination generate a rule whose premises are the input nodes to this combination of links. All premises of a rule are conjuncted. Either [35], KT [9] and Subset =-=[52]-=- are three notable rule extraction algorithms in this category. Some of the main problems of the KT and the Subset algorithms are: (i) the size of the search algorithm is O(2 l ) for a hidden/output n... |

198 | Eds.) ''The MONK's Problems: A Performance Comparison of Different Learning Algorithms
- Thrun, Mitchell, et al.
- 1991
(Show Context)
Citation Context ...alignant. These instances are divided into a training set of size 341 and a test set of size 342. Other popular data sets that have been used as benchmarks for rule extraction approaches are the Monk =-=[49]-=-, Mushroom [21] and the DNA promoter [54] data sets. All three of these data sets inputs are symbolic/discrete by nature. Since we want to test more general problems that may include continuous valued... |

197 |
Chimerge: discretization of numeric attributes
- Kerber
- 1992
(Show Context)
Citation Context ... i;l are the lower and upper boundary values of interval l of input X i respectively. Different discretization approaches can be exploited to compute discretization boundaries of input features X i s =-=[46, 5, 23, 26, 56]-=-. Full-RE uses the Chi2 [25] algorithm 2 , a powerful discretization tool, to compute discretization boundaries of input features. When Full-RE finds more 2 We are thankful to Liu and Setiono for maki... |

194 | Refinement of approximate domain theories by knowledge-based neural networks
- Towell, Shavlik, et al.
- 1990
(Show Context)
Citation Context ... : : : : : : : : : : : : : : : : : 31 6 Conclusionss1 INTRODUCTION 3 1 Introduction Several researchers have investigated the design of hybrid systems that combine expert and connectionist subsystems =-=[44, 45, 54, 10, 16, 15, 27]-=-. The typical result is a Knowledge Based Neural Network (KBNN) system with four phases: (i) the rule base representation phase, where initial domain knowledge is extracted and represented in a symbol... |

189 |
On changing continuous attributes into ordered discrete attributes
- Catlett
- 1991
(Show Context)
Citation Context ... i;l are the lower and upper boundary values of interval l of input X i respectively. Different discretization approaches can be exploited to compute discretization boundaries of input features X i s =-=[46, 5, 23, 26, 56]-=-. Full-RE uses the Chi2 [25] algorithm 2 , a powerful discretization tool, to compute discretization boundaries of input features. When Full-RE finds more 2 We are thankful to Liu and Setiono for maki... |

185 |
Learning and Extracting Finite State Automata with Second-Order Recurrent Neural Networks
- Giles, Miller, et al.
- 1992
(Show Context)
Citation Context ... functions and explaining the trained neural network [17, 48, 47, 24] 2.2.4 Extracting rules from recurrent networks Recurrent networks have shown great success in representing finite state languages =-=[14, 55]-=- and deterministic finite state automata [13]. Omlin and Giles, [33] have developed a heuristic algorithm to extract grammar rules in the form of Deterministic Finite-state Automata (DFA) from discret... |

183 | Knowledge-based artificial neural networks
- Towell, Shavlik
- 1994
(Show Context)
Citation Context ... system to provide explanation power. KBNNs attempt to exploit the complementary properties of knowledge based and neural network paradigms to obtain more powerful and robust systems. HIA [44], KBANN =-=[53, 34]-=-, RAPTURE [27] and KBCNN [10, 11] are examples of KBNN hybrid systems. Figure 1 sketches typical components of a KBNN system that combines rule-based and connectionist paradigms. Researchers have also... |

157 |
Neural-Network-Based Fuzzy Logic Control and Decision System”, in
- LIN, LEE
- 1991
(Show Context)
Citation Context ...eural Networks (FLNN or NeuroFuzzy) hybrid systems. In FLNNs, the neural network subsystem is typically used to adapt membership functions of fuzzy variables [6], or to refine and extract fuzzy rules =-=[48, 47, 24]-=-. Rules Extracted Updated Rule-Based System Output Hybrid Rule-based Decisions Concepts Learned NN (MLP) Architecture Connectionist Dec isions Decisions Integrated Decision Maker Training Examples Rul... |

155 |
Cancer Diagnosis via Linear Programming
- Mangasarian, Wolberg
- 1990
(Show Context)
Citation Context ...ut pattern has four continuous input features: I 1 = Sepal-length , I 2 = Sepal-width, I 3 = Petal-length, and I 4 =Petal-width. 3. Breast-Cancer data set which has nine inputs and two output classes =-=[28, 32]-=-. The input features are: X 1 = Clump Thickness, X 2 = Uniformity of Cell Size, X 3 = Uniformity of Cell Shape, X 4 = Marginal Adhesion, X 5 = Single Epithelial Cell Size, X 6 = Bare Nuclei, X 7 = Bla... |

147 | Chi2: Feature selection and discretization of numeric attributes
- Liu, Setiono
- 1995
(Show Context)
Citation Context ...s of interval l of input X i respectively. Different discretization approaches can be exploited to compute discretization boundaries of input features X i s [46, 5, 23, 26, 56]. Full-RE uses the Chi2 =-=[25]-=- algorithm 2 , a powerful discretization tool, to compute discretization boundaries of input features. When Full-RE finds more 2 We are thankful to Liu and Setiono for making their Chi2 source code av... |

139 |
Changing the rules: a comprehensive approach to theory refinement
- Ourston, Mooney
- 1990
(Show Context)
Citation Context ...exceeds the bias of the current node. (ii) For each of these combination generate a rule whose premises are the input nodes to this combination of links. All premises of a rule are conjuncted. Either =-=[35]-=-, KT [9] and Subset [52] are three notable rule extraction algorithms in this category. Some of the main problems of the KT and the Subset algorithms are: (i) the size of the search algorithm is O(2 l... |

133 |
Introduction to probability and statistics
- Mendenhall
- 1998
(Show Context)
Citation Context ...dard deviation of input feature X i. In (2), s i is multiplied by “2” to provide a wider distribution of input X i (a range of m i ± 2s i will contain approximately 95 percent of the X i measurements =-=[23]-=-, if X i is normally distributed). If more detailed rules are required (i.e., the comprehensibility measure p > 1), then Partial-RE starts looking for combinations of two unmarked links starting from ... |

107 |
On fuzzy modeling using fuzzy neural networks with back-propagation algorithm
- Horikawa, Furahashi, et al.
- 1992
(Show Context)
Citation Context ...ncerned with combining neural networks and fuzzy logic. Some FLNN systems include a fuzzy rule extraction module for refining fuzzy sets membership functions and explaining the trained neural network =-=[17, 48, 47, 24]-=- 2.2.4 Extracting rules from recurrent networks Recurrent networks have shown great success in representing finite state languages [14, 55] and deterministic finite state automata [13]. Omlin and Gile... |

102 |
The truth will come to light: Directions and challenges in extracting the knowledge embedded within trained artificial neural networks
- Tickle, Andrews, et al.
- 1998
(Show Context)
Citation Context ...eural networks, rule evaluation. 1 INTRODUCTION S EVERAL researchers have investigated the design of hybrid systems that combine expert and connectionist subsystems [40], [47], [7], [12], [11], [20], =-=[48]-=-. The typical result of a transformational type of hybridization [49] is a Knowledge-Based Neural Network (KBNN) system with theory refinement capabilities, usually involving four phases: 1) the rule-... |

86 |
Induction of Finite-State Languages Using Second-Order Recurrent
- Watrous, Kuhn
- 1992
(Show Context)
Citation Context ... functions and explaining the trained neural network [17, 48, 47, 24] 2.2.4 Extracting rules from recurrent networks Recurrent networks have shown great success in representing finite state languages =-=[14, 55]-=- and deterministic finite state automata [13]. Omlin and Giles, [33] have developed a heuristic algorithm to extract grammar rules in the form of Deterministic Finite-state Automata (DFA) from discret... |

83 | Using sampling and queries to extract rules from trained neural networks
- Craven, Shavlik
- 1994
(Show Context)
Citation Context ... rule extraction approach 2 RULE EXTRACTION 8 is the algorithm developed by Saito and Nakano to extract medical diagnostic rules from a trained network [39]. BRAINNE [40], Rule-extraction-as-learning =-=[7]-=-, and DEDEC [50] are other examples of extracting rules by investigating the input-output mapping of a trained network. In this paper we refer to this class as the Black-box Rule Extraction (BRE) cate... |

83 |
Concept Acquisition Through Representational Adjustment
- Schlimmer
- 1987
(Show Context)
Citation Context ... instances are divided into a training set of size 341 and a test set of size 342. Other popular data sets that have been used as benchmarks for rule extraction approaches are the Monk [49], Mushroom =-=[21]-=- and the DNA promoter [54] data sets. All three of these data sets inputs are symbolic/discrete by nature. Since we want to test more general problems that may include continuous valued variables, Iri... |

80 |
Rule learning by searching on adapted nets
- Fu
- 1991
(Show Context)
Citation Context ...he bias of the current node. (ii) For each of these combination generate a rule whose premises are the input nodes to this combination of links. All premises of a rule are conjuncted. Either [35], KT =-=[9]-=- and Subset [52] are three notable rule extraction algorithms in this category. Some of the main problems of the KT and the Subset algorithms are: (i) the size of the search algorithm is O(2 l ) for a... |

68 | Extraction of Rules from Discrete-Time Recurrent Neural Networks
- Omlin, Giles
- 1992
(Show Context)
Citation Context ...2.4 Extracting rules from recurrent networks Recurrent networks have shown great success in representing finite state languages [14, 55] and deterministic finite state automata [13]. Omlin and Giles, =-=[33]-=- have developed a heuristic algorithm to extract grammar rules in the form of Deterministic Finite-state Automata (DFA) from discrete-time neural networks and specifically from second-order networks. ... |

65 |
NN-driven fuzzy reasoning
- Takagi, Hayashi
- 1991
(Show Context)
Citation Context ...eural Networks (FLNN or NeuroFuzzy) hybrid systems. In FLNNs, the neural network subsystem is typically used to adapt membership functions of fuzzy variables [6], or to refine and extract fuzzy rules =-=[48, 47, 24]-=-. Rules Extracted Updated Rule-Based System Output Hybrid Rule-based Decisions Concepts Learned NN (MLP) Architecture Connectionist Dec isions Decisions Integrated Decision Maker Training Examples Rul... |

55 |
Neural networks in computer intelligence
- Fu
- 1994
(Show Context)
Citation Context ...wer. KBNNs attempt to exploit the complementary properties of knowledge based and neural network paradigms to obtain more powerful and robust systems. HIA [44], KBANN [53, 34], RAPTURE [27] and KBCNN =-=[10, 11]-=- are examples of KBNN hybrid systems. Figure 1 sketches typical components of a KBNN system that combines rule-based and connectionist paradigms. Researchers have also combined connectionist systems w... |

55 | Symbolic representation of neural network
- Setiono, Liu
- 1996
(Show Context)
Citation Context ...ms like "promoter recognition in DNA nucleotides" for which it is a natural fit [52]. NeuroRule is another rule extraction approach that uses different combinations of weighted links to extr=-=act rules [43]-=-. The main difference between NeuroRule and MofN is that the former extracts rules from networks after pruning their architectures and then discretizing their hidden 2 RULE EXTRACTION 7 units activati... |

45 | Medical diagnostic expert system based on PDP models”[J - Saito, Nakano |

40 |
Marker-passing over microfeatures: Towards a hybrid symbolic/connectionist model
- Hendler
- 1989
(Show Context)
Citation Context ... : : : : : : : : : : : : : : : : : 31 6 Conclusionss1 INTRODUCTION 3 1 Introduction Several researchers have investigated the design of hybrid systems that combine expert and connectionist subsystems =-=[44, 45, 54, 10, 16, 15, 27]-=-. The typical result is a Knowledge Based Neural Network (KBNN) system with four phases: (i) the rule base representation phase, where initial domain knowledge is extracted and represented in a symbol... |

38 |
Knowledge-based connectionism for revising domain theories
- Fu
- 1993
(Show Context)
Citation Context ... : : : : : : : : : : : : : : : : : 31 6 Conclusionss1 INTRODUCTION 3 1 Introduction Several researchers have investigated the design of hybrid systems that combine expert and connectionist subsystems =-=[44, 45, 54, 10, 16, 15, 27]-=-. The typical result is a Knowledge Based Neural Network (KBNN) system with four phases: (i) the rule base representation phase, where initial domain knowledge is extracted and represented in a symbol... |

38 | Hybrid Neural Systems: From Simple Coupling to Fully
- McGarry, Wermter, et al.
- 1999
(Show Context)
Citation Context ...have investigated the design of hybrid systems that combine expert and connectionist subsystems [40], [47], [7], [12], [11], [20], [48]. The typical result of a transformational type of hybridization =-=[49]-=- is a Knowledge-Based Neural Network (KBNN) system with theory refinement capabilities, usually involving four phases: 1) the rule-base representation phase, where initial domain knowledge is extracte... |

33 | Structural adaptation and generalization in supervised feedforward networks
- Ghosh, Tumer
- 1994
(Show Context)
Citation Context ...rk is trained using the backpropagation algorithm with momentum as well as a regularization term P which adds \Gamma2w jk w 0 2 w 0 2 +w 2 jk to the weight update term in the backpropagation equation =-=[12]-=-. Cross validation is used for the stopping criteria. 2. Network architectures and data reduction: for the iris problem, an MLP with 4 input, 6 hidden, and 3 output nodes is used for the three experim... |

33 | Understanding Neural Networks via Rule Extraction
- Setiono, Liu
(Show Context)
Citation Context ...by both NeuroRule and C4.5rules algorithms [43]. The main reason of choosing NeuroRule and C4.5rules is that they have previously been used to extract rules for the same two databases used by Full-RE =-=[42]-=-. Moreover, they both extract comprehensive rules with relatively high correct classification rate as reported by the authors of NeuroRule [42]. For iris problem, we also compare the set of rules extr... |

31 | Extracting Rules from Pruned Neural Networks for Breast Cancer Diagnosis
- Setiono
- 1996
(Show Context)
Citation Context ...put, 6 hidden, and 2 output nodes) are presented in Table 7. The rules extracted by NeuroRule from the best among the pruned 100 MLP network architectures (6 inputs, 1 hidden, and 2 output nodes) are =-=[43, 41]-=-: Rule 1: If X 1 ! 7:0 and X 2 ! 8:0 and X 3 ! 3:0 and X 8 ! 9:0, then Benign Rule 2: If X 1 ! 7:0 and X 2 ! 8:0 and X 3 ! 3:0 and X 6 ! 9:0, then Benign Rule 3: If X 2 ! 8:0 and X 3 ! 3:0 and X 6 ! 3... |

29 | Combining connectionist and symbolic learning to re ne certainty-factor rule-bases
- Mahoney, Mooney
- 1993
(Show Context)
Citation Context |

28 | The Connectionist Scientist Game: Rule Extraction and Refi nement in a Neural Network
- McMillan, Mozer, et al.
- 1991
(Show Context)
Citation Context ...n approaches that extract rules from feedforward ANNs have been reported. The main difference between them and the approaches mentioned above is that they extract rules from specialized ANNs. RuleNet =-=[30]-=- and RULEX [3, 2] are two examples of this class of approaches. RULEX extracts rules from a Constrained Error Back-Propagation (CEBP) MLP network, similar to Radial Basis Function (RBF) networks. Each... |

24 | Heuristically expanding knowledge-based neural networks
- Opitz, Shavlik
- 1993
(Show Context)
Citation Context ... system to provide explanation power. KBNNs attempt to exploit the complementary properties of knowledge based and neural network paradigms to obtain more powerful and robust systems. HIA [44], KBANN =-=[53, 34]-=-, RAPTURE [27] and KBCNN [10, 11] are examples of KBNN hybrid systems. Figure 1 sketches typical components of a KBNN system that combines rule-based and connectionist paradigms. Researchers have also... |

22 | Learning a Class of Large Finite State Machines With a Recurrent Neural Network
- Giles, Home, et al.
- 1995
(Show Context)
Citation Context ...ork [17, 48, 47, 24] 2.2.4 Extracting rules from recurrent networks Recurrent networks have shown great success in representing finite state languages [14, 55] and deterministic finite state automata =-=[13]-=-. Omlin and Giles, [33] have developed a heuristic algorithm to extract grammar rules in the form of Deterministic Finite-state Automata (DFA) from discrete-time neural networks and specifically from ... |

21 |
Geva S.; Rule extraction from a constrained error back propagation
- Andrews
- 1994
(Show Context)
Citation Context ...at extract rules from feedforward ANNs have been reported. The main difference between them and the approaches mentioned above is that they extract rules from specialized ANNs. RuleNet [30] and RULEX =-=[3, 2]-=- are two examples of this class of approaches. RULEX extracts rules from a Constrained Error Back-Propagation (CEBP) MLP network, similar to Radial Basis Function (RBF) networks. Each hidden node in t... |

18 |
ESPRESSO-MV: Algorithms for multiple-valued logic minimization
- Rudell, Sangiovann-Vincentelli
- 1985
(Show Context)
Citation Context ...f available logic minimization tools. 2. Extracted rules are optimal and cannot be simplified any further. Hence, no rewriting procedure is required. 1 Espresso is a software package for logic design =-=[38]-=-. 3 PROPOSED RULE EXTRACTION APPROACHES 11 3. The extracted rules do not depend on the number of layers of the trained network. 4. The set of rules extracted by BIO-RE is comprehensive and understanda... |

15 | Theory Refinement for Bayesian Networks with Hidden Variables
- Ramachandran, Mooney
- 1998
(Show Context)
Citation Context ...e and extract fuzzy rules [42], [41], [18]. Neural networks have also been used for refinement of initial theories expressed in other knowledge representation schemes such as Bayesian Belief networks =-=[29]-=-. Extraction of symbolic rules from trained Artificial Neural Networks (ANNs) is an important feature of comprehensive hybrid systems, as it helps to: 1) Alleviate the knowledge acquisition problem an... |

14 |
DEDEC: decision detection by rule extraction from neural networks
- Tickle, Orlowski, et al.
- 1995
(Show Context)
Citation Context ...n approach 2 RULE EXTRACTION 8 is the algorithm developed by Saito and Nakano to extract medical diagnostic rules from a trained network [39]. BRAINNE [40], Rule-extraction-as-learning [7], and DEDEC =-=[50]-=- are other examples of extracting rules by investigating the input-output mapping of a trained network. In this paper we refer to this class as the Black-box Rule Extraction (BRE) category because rul... |

14 |
A Methodology for Extracting Rules from Trained
- Tickle, Orlowski, et al.
- 1996
(Show Context)
Citation Context ...ss of rule extraction techniques is 00 pedagogical 00 approaches [3]. For example, DEDEC extracts rules by ranking the inputs of an ANN according to their importance (contribution) to the ANN outputs =-=[51]-=-. This ranking process is done by examining the weight vectors of the ANN, which puts DEDEC on the border between LRE and BRE techniques. The next step in DEDEC is to cluster these ranked inputs and u... |

13 |
S.: Inserting and Extracting Knowledge from Constrained Error Back Propagation Networks
- Andrews, Geva
- 1995
(Show Context)
Citation Context ...at extract rules from feedforward ANNs have been reported. The main difference between them and the approaches mentioned above is that they extract rules from specialized ANNs. RuleNet [30] and RULEX =-=[3, 2]-=- are two examples of this class of approaches. RULEX extracts rules from a Constrained Error Back-Propagation (CEBP) MLP network, similar to Radial Basis Function (RBF) networks. Each hidden node in t... |

11 |
Understanding Catastrophic Interference in Neural Nets
- Sharkey, Sharkey
- 1994
(Show Context)
Citation Context ...l domain knowledge. 2) Provide reasoning and explanation capabilities. 3) Support cross-referencing and verification capabilities. 4) Alleviate the “catastrophic interference” problem of certain ANNs =-=[37]-=-. For models such as MLPs, 1 it has been observed that if a network originally trained on one task (data set) is subsequently trained on a different task (statistically different data set), then its p... |

8 |
Learning relations from noisy examples: An empirical comparison of LINUS and FOIL
- Dzerisko, Lavrac
- 1991
(Show Context)
Citation Context ...om the KBANN trained network through six main procedures. Rules extracted by MofN are significantly superior than rules extracted by other symbolic approaches such as C4.5 [37], Either [35] and LINUS =-=[8] at least -=-for problems like "promoter recognition in DNA nucleotides" for which it is a natural fit [52]. NeuroRule is another rule extraction approach that uses different combinations of weighted lin... |

8 |
Discretization of ordinal attributes and feature selection
- Liu, Setiono
- 1995
(Show Context)
Citation Context |

8 |
Automated Knowledge Acquisition of Rules With Continuously Valued Attributes
- Sestito, Dillon
- 1992
(Show Context)
Citation Context ...ing behavior. An example of such a rule extraction approach 2 RULE EXTRACTION 8 is the algorithm developed by Saito and Nakano to extract medical diagnostic rules from a trained network [39]. BRAINNE =-=[40]-=-, Rule-extraction-as-learning [7], and DEDEC [50] are other examples of extracting rules by investigating the input-output mapping of a trained network. In this paper we refer to this class as the Bla... |

7 | Hybrid Intelligent Architecture and Its Application to Water Reservoir Control
- Taha, Ghosh
- 1997
(Show Context)
Citation Context |

6 |
Hybrid neural network and rule-based pattern recognition system capable of self-modification
- Glover, Silliman, et al.
- 1990
(Show Context)
Citation Context |

6 | A hybrid intelligent architecture for refining input characterization and domain knowledge
- Taha, Ghosh
- 1995
(Show Context)
Citation Context |

6 |
Medical Diagnostic Expert System Based on
- Saito, Nakano
- 1988
(Show Context)
Citation Context ...by examining their inputoutput mapping behavior. An example of such a ruleextraction approach is the algorithm developed by Saito and Nakano to extract medical diagnostic rules from a trained network =-=[31]-=-. BRAINNE [33], Rule-extraction-aslearning [6], and DEDEC [44] are other examples of extracting rules by investigating the input-output mapping of a trained network. In this paper we refer to this cla... |

4 | Controlling water reservoirs using a hybrid intelligent architecture
- Taha, Ghosh
- 1995
(Show Context)
Citation Context |

2 |
A fuzzy neural hybrid system
- Challo, McLauchlan, et al.
- 1994
(Show Context)
Citation Context ...uzzy logic systems to obtain Fuzzy Logic Neural Networks (FLNN or NeuroFuzzy) hybrid systems. In FLNNs, the neural network subsystem is typically used to adapt membership functions of fuzzy variables =-=[6]-=-, or to refine and extract fuzzy rules [48, 47, 24]. Rules Extracted Updated Rule-Based System Output Hybrid Rule-based Decisions Concepts Learned NN (MLP) Architecture Connectionist Dec isions Decisi... |

2 |
Rule extraction from neural networks
- Howes, Crook
- 1996
(Show Context)
Citation Context ...chitectures and then discretizing their hidden 2 RULE EXTRACTION 7 units activation values. Recently, Howes and Crook introduced another algorithm that extracts rules from feedforward neural networks =-=[18]-=-. The network architecture used by this algorithm is restricted to one hidden layer network trained with a binary sigmoid activation function. The rationale behind this algorithm is to extract the max... |

2 |
A map method for synthesis of combinational logic circuits
- Karnough
- 1953
(Show Context)
Citation Context ...ed in the previously described binary rule format) from the truth table of step 2. Any available boolean simplification method can be used to perform step 3 of the BIO-RE algorithm (e.g. Karnough map =-=[22]-=-, algebraic manipulation, or a tabulation method [29]). We used Espresso 1 to generate the extracted rules [4]. Rules extracted by BIO-RE are represented in the format: If [Not] Input-Variable [And [N... |

2 |
Introduction to Probability and Statistics, Fifth Edition
- Mendenhall
- 1979
(Show Context)
Citation Context ...is multiplied by "2" to provide a 3 PROPOSED RULE EXTRACTION APPROACHES 12 wider distribution of input X i ( a range of �� i \Sigma 2oe i will contain approximately 95 percent of the X i=-= measurements [31]-=- if X i is normally distributed). If more detailed rules are required (i.e. the comprehensibility measure p ? 1), then PartialRE starts looking for combinations of two unmarked links starting from the... |

2 |
Medical diagnostic expert system based on DPD model
- Saito, Nakano
- 1988
(Show Context)
Citation Context ...ut-output mapping behavior. An example of such a rule extraction approach 2 RULE EXTRACTION 8 is the algorithm developed by Saito and Nakano to extract medical diagnostic rules from a trained network =-=[39]-=-. BRAINNE [40], Rule-extraction-as-learning [7], and DEDEC [50] are other examples of extracting rules by investigating the input-output mapping of a trained network. In this paper we refer to this cl... |

2 |
A generation methods for fuzzy rules using neural networks with planar lattice architecture
- Tazaki, Inoue
- 1994
(Show Context)
Citation Context ...eural Networks (FLNN or NeuroFuzzy) hybrid systems. In FLNNs, the neural network subsystem is typically used to adapt membership functions of fuzzy variables [6], or to refine and extract fuzzy rules =-=[48, 47, 24]-=-. Rules Extracted Updated Rule-Based System Output Hybrid Rule-based Decisions Concepts Learned NN (MLP) Architecture Connectionist Dec isions Decisions Integrated Decision Maker Training Examples Rul... |

2 |
et al. Extracting and learning an unknown grammar with recurrent neural networks
- Giles
- 1992
(Show Context)
Citation Context ...twork [13], [42], [41], [18]. 2.2.4 Extracting Rules from Recurrent Networks Recurrent networks have shown great success in representing finite state languages and deterministic finite state automata =-=[10]-=-. Omlin and Giles [25] have developed a heuristic algorithm to extract grammar rules in the form of Deterministic Finite-state Automata (DFA) from discrete-timesTAHA AND GHOSH: SYMBOLIC INTERPRETATION... |

2 | Rule set quality measures for inductive learning algorithms
- Clair
- 1996
(Show Context)
Citation Context ...nds on the application domain. For example, in image understanding, one may be primarily interested in finding relational visual structures [4]. Also, one can define metrics for the complete rule set =-=[17]-=- or for individual rules. We will do the former in Section 5. In this section, we evaluate individual rules based on the following primary goals. 1) Find the order of the extracted rules that maximize... |

1 |
The role of machine learning in building image interpretation systems
- Caelli, Bischof
- 1996
(Show Context)
Citation Context ... Evaluation The suitability of rule evaluation criteria depends on the application domain. For example, in image understanding, one may be primarily interested in finding relational visual structures =-=[4]-=-. Also, one can define metrics for the complete rule set [17] or for individual rules. We will do the former in Section 5. In this section, we evaluate individual rules based on the following primary ... |

1 |
A hybrid intelligent architecture for revising domain knowledge
- Taha
- 1997
(Show Context)
Citation Context ...ustable parameters D and p (which determines the number of premises in a rule) provides flexibility to the Partial-RE algorithm. Several other differences in implementational details are described in =-=[38]-=-. Partial-RE is easily parallelizable, as nodes can be inspected concurrently. Experimental results show that Partial-RE algorithm is suitable for large size problems, since extracting all possible ru... |

1 | Theory Refinement of Bayesian Netowrks with Hidden Variables - Ramachandran - 1997 |