Search

Applying machine learning techniques to spatial-temporal data poses the question that how the recorded location and time for training samples should contribute to the training and testing process. The prior knowledge of how spatial-temporal phenomena are autocorrelated cannot be properly captured by machine learning techniques, which either ignore location and time altogether or consider them as input features. Not to mention that the latter approach leads to slightly increased sparseness of data in the feature space and more free parameters in the predictor; thus, demanding for larger training datasets. We use the prior knowledge about the spatial-temporal autocorrelation to determine how relevant each training sample would be, given its spatial and temporal distances to the irresponsive (unlabeled) sample. Weighted machine learning techniques use this prior knowledge by taking the relevance of training samples with regard to the irresponsive sample into account as training samples’ weights. The proposed approach overcomes the aforementioned issues by enriching the training process with the prior knowledge about spatial-temporal autocorrelation. Because the spatial-temporal weight of training samples depends on the irresponsive sample's location and time, the machine needs to be trained separately for each irresponsive sample. However, we show that in practice using only a small subset of training samples with largest spatial-temporal weights not only mitigates the training time but also results in the best accuracy in most cases.

A teacher should arguably know about learning theories (LTs) in order to make daily pedagogical decisions. However, little literature exists on the role of LTs in teacher education. Eight Dutch teacher educators were interviewed on LTs in their curriculum. LTs were unanimously considered important but huge variation was found in what and how LTs are taught. Several functions of LTs were mentioned, with underpinning of pedagogical decisions using LT considered to be the essence of higher education. However, respondents doubted whether this is ever achieved. This suggests an additional paradigm – reality gap in teacher education.

Graphs are a natural choice to encode data in many real–world applications. In fact, a graph can describe a given pattern as a complex structure made up of parts (the nodes) and relationships between them (the edges). Despite their rich representational power, most of machine learning approaches cannot deal directly with inputs encoded by graphs. Indeed, Graph Neural Networks (GNNs) have been devised as an extension of recursive models, able to process general graphs, possibly undirected and cyclic. In particular, GNNs can be trained to approximate all the “practically useful” functions on the graph space, based on the classical inductive learning approach, realized within the supervised framework. However, the information encoded in the edges can actually be used in a more refined way, to switch from inductive to transductive learning. In this paper, we present an inductive–transductive learning scheme based on GNNs. The proposed approach is evaluated both on artificial and real–world datasets showing promising results. The recently released GNN software, based on the Tensorflow library, is made available for interested users.

[EN] A common way of learning to perform a task is to observe how it is carried out by experts. However, it is well known that for most tasks there is no unique way to perform them. This is especially noticeable the more complex the task is because factors such as the skill or the know-how of the expert may well affect the way she solves the task. In addition, learning from experts also suffers of having a small set of training examples generally coming from several experts (since experts are usually a limited and ex- pensive resource), being all of them positive examples (i.e. examples that represent successful executions of the task). Traditional machine learning techniques are not useful in such scenarios, as they require extensive training data. Starting from very few executions of the task presented as activity sequences, we introduce a novel inductive approach for learning multiple models, with each one representing an alter- native strategy of performing a task. By an iterative process based on generalisation and specialisation, we learn the underlying patterns that capture the different styles of performing a task exhibited by the examples. We illustrate our approach on two common activity recognition tasks: a surgical skills training task and a cooking domain. We evaluate the inferred models with respect to two metrics that measure how well the models represent the examples and capture the different forms of executing a task showed by the examples. We compare our results with the traditional process mining approach and show that a small set of meaningful examples is enough to obtain patterns that capture the different strategies that are followed to solve the tasks., This work has been partially supported by the EU (FEDER) and the Spanish MINECO under grants TIN2014-61716-EXP (SUPERVASION) and RTI2018-094403-B-C32, and by Generalitat Valenciana under grant PROMETEO/2019/098. David Nieves is also supported by the Spanish MINECO under FPI grant (BES-2016-078863).

Mixing examples of different categories (interleaving) has been shown to promote inductive learning as compared with presenting examples of the same category together (massing). In three studies, we tested whether the advantage of interleaving is exclusively due to the mixing of examples from different categories or to the temporal gap introduced between presentations. In addition, we also tested the role of working memory capacity (WMC). Results showed that the mixing of examples might be the key component that determines improved induction. WMC might also be involved in the interleaving effect: participants with high spans seemed to profit more than participants with low spans from interleaved presentations. Our findings have relevant implications for education. Practice schedules should be individually customised so society as a whole can profit from differences between learners. (PsycINFO Database Record (c) 2017 APA, all rights reserved)

State of the art algorithms for many pattern recognition problems rely on data-driven deep network models. Training these models requires a large labeled dataset and considerable computational resources. Also, it is difficult to understand the working of these learned models, limiting their use in some critical applications. Toward addressing these limitations, our architecture draws inspiration from research in cognitive systems, and integrates the principles of commonsense logical reasoning, inductive learning, and deep learning. As a motivating example of a task that requires explainable reasoning and learning, we consider Visual Question Answering in which, given an image of a scene, the objective is to answer explanatory questions about objects in the scene, their relationships, or the outcome of executing actions on these objects. In this context, our architecture uses deep networks for extracting features from images and for generating answers to queries. Between these deep networks, it embeds components for non-monotonic logical reasoning with incomplete commonsense domain knowledge, and for decision tree induction. It also incrementally learns and reasons with previously unknown constraints governing the domain's states. We evaluated the architecture in the context of datasets of simulated and real-world images, and a simulated robot computing, executing, and providing explanatory descriptions of plans and experiences during plan execution. Experimental results indicate that in comparison with an “end to end” architecture of deep networks, our architecture provides better accuracy on classification problems when the training dataset is small, comparable accuracy with larger datasets, and more accurate answers to explanatory questions. Furthermore, incremental acquisition of previously unknown constraints improves the ability to answer explanatory questions, and extending non-monotonic logical reasoning to support planning and diagnostics improves the reliability and efficiency of computing and executing plans on a simulated robot.

We present Sigma*, a novel technique for learning symbolic models of software behavior. Sigma* addresses the challenge of synthesizing models of software by using symbolic conjectures and abstraction. By combining dynamic symbolic execution to discover symbolic input-output steps of the programs and counterexample guided abstraction refinement to over-approximate program behavior, Sigma* transforms arbitrary source representation of programs into faithful input-output models. We define a class of stream filters---programs that process streams of data items---for which Sigma* converges to a complete model if abstraction refinement eventually builds up a sufficiently strong abstraction. In other words, Sigma* is complete relative to abstraction. To represent inferred symbolic models, we use a variant of symbolic transducers that can be effectively composed and equivalence checked. Thus, Sigma* enables fully automatic analysis of behavioral properties such as commutativity, reversibility and idempotence, which is useful for web sanitizer verification and stream programs compiler optimizations, as we show experimentally. We also show how models inferred by Sigma* can boost performance of stream programs by parallelized code generation.

Traditional supervised approaches realize an inductive learning process: A model is learnt from labeled examples, in order to predict the labels of unseen examples. On the other hand, transductive learning is less ambitious. It can be thought as a procedure to learn the labels on a training set, while, simultaneously, trying to guess the best labels on the test set. Intuitively, transductive learning has the advantage of being able to directly use training patterns while deciding on a test pattern. Thus, transductive learning faces a simpler problem with respect to inductive learning. In this paper, we propose a preliminary comparative study between a simple transductive model and a pure inductive model, where the learning architectures are based on feedforward neural networks. The goal is to understand how transductive learning affects the complexity measured by the number of hidden neurons of the exploited neural networks. Preliminary experimental results are reported on the classical two spirals problem.

This paper describes RULES3-EXT, a new algorithm for inductive learning. It has been developed to cope with some drawbacks of RULES-3 induction algorithm. The extra features of RULES3-EXT are (1) The number of required files to extract a knowledge base (a set of rules) is reduced to 2 from 3 (2) The repeated examples are eliminated, (3) The users are able to change the order of attributes and (4) The system is able to fire rule(s) partially if any of the extracted rules cannot fully be satisfied by an unseen example. The new algorithm has been tested on well known data sets and the efficiency found to be superior to that of RULES-3.

This paper presents CSLS, a symbiotic combination of inductive and neural learning. CSLS has two components, an induction algorithm to carry out inductive learning and a multi-layer perceptron (MLP) to implement neural learning. The paper outlines the operation of the components of CSLS and describes how the combined system is designed to utilise the individual strengths of inductive and neural learning to the best advantage. The paper gives the results of evaluating CSLS on the IRIS data and Breast-Cancer-Wisconsin-data classification problems. These clearly demonstrate the main benefit of the symbiotic combination: the combined system performs better than either of its components.

In recent years, there has been a growing amount of research on inductive learning. Out of this research a number of promising algorithms have surfaced. In the paper after a brief description of knowledge acquisition, induction and inductive learning; RULES family of inductive learning algorithms, their strengths as well as weaknesses are explained and discussed. The applications of inductive learning and particularly the applications of RULES family of algorithms are overviewed.

Due to the daily increase in the size of data, machine learning has become a popular approach for intelligent processing of data. In particular, machine learning algorithms are used to discover meaningful knowledge or build predictive models from data. For example, inductive learning algorithms involve generation of rules which can be in the form of either a decision tree or if-then rules. However, most of learning algorithms suffer from overfitting of training data. In other words, these learning algorithms can build models that perform extremely well on training data but poorly on other data. The overfitting problem is originating from both learning algorithms and data. In this context, the nature of machine learning problem can be referred to as bias and variance. The former is originating from learning algorithms whereas the latter is originating from data. Therefore, the reduction of overfitting can be achieved through scaling up algorithms on one side or scaling down data on the other side. Both bias and variance can be reduced through use of ensemble learning approaches. This paper introduces particular ways to address the issues on overfitting of rule based classifiers through both scaling up algorithms and scaling down data in the context of ensemble learning.

This paper focuses on coordinated inductive learning, concerning how agents with inductive learning capabilities can coordinate their learnt hypotheses with other agents. Coordination in this context means that the hypothesis learnt by one agent is consistent with the data known to the other agents. In order to address this problem, we present A-MAIL, an argumentation approach for agents to argue about hypotheses learnt by induction. A-MAIL integrates, in a single framework, the capabilities of learning from experience, communication, hypothesis revision and argumentation. Therefore, the A-MAIL approach is one step further in achieving autonomous agents with learning capabilities which can use, communicate and reason about the knowledge they learn from examples. © 2014, The Author(s)., Research partially funded by the projects Next-CBR (TIN2009-13692-C03-01) and Cognitio (TIN2012-38450- C03-03) [both co-funded with FEDER], Agreement Technologies (CONSOLIDER CSD2007-0022), and by the Grants 2009-SGR-1433 and 2009-SGR-1434 of the Generalitat de Catalunya.

The growing amount of information in the world has increased the need for com- puterized classification of different objects. This situation is present in higher education as well where the possibility of effortless detection of similarity between different study courses would give the opportunity to organize student exchange programmes effectively and facilitate curriculum management and development. This area which currently relies on manual time- consuming expert activities could benefit from application of smartly adapted machine learn - ing technologies. Data in this problem domain is complex leading to inability for automatic classification approaches to always reach the desired result in terms of classification accuracy. Therefore, our approach suggests an automated/semi-automated classification solution, which incorporates both machine learning facilities and interactive involvement of a domain expert for improving classification results. The system's prototype has been implemented and experi - ments are carried out. This interactive classification system allows to classify educational data, which often comes in unstructured or semi-structured, incomplete and/or insufficient form, thus reducing the number of misclassified instances significantly in comparison with the automatic machine learning approach.

Inductive inference can be considered as one of the fundamental paradigms of algorithmic learning theory. We survey results recently obtained and show their impact to potential applications. Since the main focus is put on the efficiency of learning, we also deal with postulates of naturalness and their impact to the efficiency of limit learners. In particular, we look at the learnability of the class of all pattern languages and ask whether or not one can design a learner within the paradigm of learning in the limit that is nevertheless efficient. For achieving this goal, we deal with iterative learning and its interplay with the hypothesis spaces allowed. This interplay has also a severe impact to postulates of naturalness satisfiable by any learner. Furthermore, since a limit learner is only supposed to converge, one never knows at any particular learning stage whether or not the learner did already succeed. The resulting uncertainty may be prohibitive in many applications. We survey results to resolve this problem by outlining a new learning model, called stochastic finite learning. Though pattern languages can neither be finitely inferred from positive data nor PAC-learned, our approach can be extended to a stochastic finite learner that exactly infers all pattern languages from positive data with high confidence. Finally, we apply the techniques developed to the problem of learning conjunctive concepts.

In metal cutting, the problems need to be well analyzed in order to take precautions before any unexpected results are encountered. This process plays a significant role in achieving consistent quality and in controlling the overall cost of manufacturing. However, it is a difficult task that needs an expert who has a great deal of information and experience in metal cutting. In the present paper, a knowledge-based expert system (COROSolve) that investigates problems that are encountered in three main metal cutting areas: turning, milling and drilling is developed. A great deal of metal cutting operations such as external/internal turning with negative/positive inserts, aluminum turning, parting bars/tubes, grooving, profiling, recessing and threading operations in turning; face milling, square shoulder milling, end milling, multi-purpose milling and side and face milling operations in milling; and drilling operations that use solid drills or drills with indexable inserts in drilling are taken into consideration. COROSolve gives recommendations for the cutting data (i.e., cutting speed, depth of cut, and feed) and updates the problem, cause and remedy database, thus the number of problems that the system can handle is increased.

In this paper we introduce precedence–inclusion patterns, which are sets with a strictly partially ordered set of strict partial orders, along with some additional structure. The definition of these structures reflects how multiple partial orders interact in a number of situations such as in text, images, and video. In particular, precedence–inclusion patterns generalize constituent structure trees familiar to computational linguists. Our interest in these objects was initially sparked by their connection with problems of relational learning. We develop the general mathematical theory of precedence–inclusion patterns, we show that each finite set of finite precedence–inclusion relations has a minimal most specific generalization that is unique up to isomorphism, and we explain how this result relates to relational learning.

One important disadvantage of decision tree based inductive learning algorithms is that they use some irrelevant values to establish the decision tree. This causes the final rule set to be less general. To overcome with this problem the tree has to be pruned. In this article using the recently developed RULES inductive learning algorithm, pruning of a decision tree is explained. The decision tree is extracted for an example problem using the ID3 algorithm and then is pruned using RULES. The results obtained before and after pruning are compared. This shows that the pruned decision tree is more general.

We describe an approach to machine learning from numerical data that combines both qualitative and numerical learning. This approach is carried out in two stages: (1) induction of a qualitative model from numerical examples of the behaviour of a physical system, and (2) induction of a numerical regression function that both respects the qualitative constraints and fits the training data numerically. We call this approach Q2 learning, which stands for Qualitatively faithful Quantitative learning. Induced numerical models are “qualitatively faithful” in the sense that they respect qualitative trends in the learning data. Advantages of Q2 learning are that the induced qualitative model enables a (possibly causal) explanation of relations among the variables in the modelled system, and that numerical predictions are guaranteed to be qualitatively consistent with the qualitative model which alleviates the interpretation of the predictions. Moreover, as we show experimentally the qualitative model's guidance of the quantitative modelling process leads to predictions that may be considerably more accurate than those obtained by state-of-the-art numerical learning methods. The experiments include an application of Q2 learning to the identification of a car wheel suspension system—a complex, industrially relevant mechanical system.

Inductive learning takes place when people learn a new concept or category by observing a variety of exemplars. Kornell and Bjork (2008) asked participants to learn new painting styles either by presenting different paintings of the same artist consecutively (massed presentation) or by mixing paintings of different artists (spaced presentation). In their second experiment, Kornell and Bjork (2008) showed with a final style recognition test, that spacing resulted in better inductive learning than massing. Also, by using this style recognition test, they ruled out the possibility that spacing merely resulted in a better memory for the labels of the newly learned painting styles. The findings from Kornell and Bjork’s (2008) second experiment are important because they show that the benefit of spaced learning generalizes to complex learning tasks and outcomes, and that it is not confined to rote memory learning. However, the findings from Kornell and Bjork’s (2008) second experiment have never been replicated. In the present study we performed an exact and high-powered replication of Kornell and Bjork’s (2008) second experiment with a Web-based sample. Such a replication contributes to establish the reliability of the original finding and hence to more conclusive evidence of the spacing effect in inductive learning. The findings from the present replication attempt revealed a medium-sized advantage of spacing over massing in inductive learning, which was comparable to the original effect in the experiment by Kornell and Bjork (2008). Also, the 95% confidence intervals (CI) of the effect sizes from both experiments overlapped considerably. Hence, the findings from the present replication experiment and the original experiment clearly reinforce each other.

A situated cognition theory is presented in which situated learning is viewed from a Vygotskian perspective: the developmental situativity theory. It is assumed that learning is basically a non-symbolic inductive process that is greatly enhanced by formal systems which act as tools. Instruction should provide students with an environment containing elements to allow for situated activity as well as with appropriate formal tools. Tarski's World, which is an instructional Interactive Graphical Representation System in formal logic is an example of such instruction. The developmental situativity theory was supported empirically: Tarski's World was found to facilitate conditional reasoning, in contrast to instruction that is based on only-situated or only-formal tasks.

One method for reducing the time required for plan generation is to learn search control rules from experience. The most common approach to learning search control knowledge has been explanationbased learning. An alternative approach is to use inductive learning. An inductive approach does not require a complete and tractable domain theory and has the potential to create more effective rules by learning from more than one example at a time. In this paper we describe Grasshopper, an inductive system that learns search control rules for a classical plan generation system. We also provide an empirical evaluation of Grasshopper by comparing it with an existing explanation-based learning system.

In this pa per, a ne\v technique to recognize and process Barcodes is introduced. The technique employs Inductive Learning. It is suitable to use, for example, in a factory to control the \\'Orkers, staff, stock ete. In this technique only vert.ical lines are considered \\·hile the spaces in bet,veen are ignored. This results faster processing. Each Barcode is considered to represent an iteın. For each Barcode a rule is ex1.racted from the necessary information using Inductive Learning. So the unnecessary information is eliminated. This causes faster processing time and less amount of memory. In order to use this technique no special hard\:vare is required. Only a PC and a Barcode reader _is enough

International audience; In reactive manufacturing control, the allocation of resources for tasks is achieved in real time. When a resource becomes available it chooses one of the tasks in its queue. This choice is made according to priority rules which are designed to optimize costs, time, etc. In this paper, the aim is to exploit a Job Shop scheduling log and simulations in order to extract knowledge enabling one to create rules for the selection of priority rules. These rules are implemented in a CASI cellular automaton. Firstly, symbolic modelling of the scheduling process is exploited to generate a decision tree from the log and simulations. Secondly, decision rules are extracted to select priority rules for execution in a specific situation. Finally, the rules are integrated in CASI which implements the decisional module of agents in a distributed manufacturing control system.

Double-blind randomized controlled trials (RCTs) are viewed as the golden standard of drug research in Western medicine. However, RCTs are far from "golden" in many respects. They are impractical for many therapies, such as for surgeries and complex lifestyle changes. They encourage a one-size-fits-all approach to medical treatment that fails to address the huge diversity among individual patients in terms of their physical and emotional symptoms, social and cultural upbringing, and other factors. Perhaps, more importantly, they do not help doctors make the best medical decisions required to produce optimal patient outcomes. To guide a search for an alternate model of medical research, three principles based on Ayurveda, an ancient and powerful system of health care that has stood the test of time, are presented. These principles, arrived at after mining Ayurvedic epistemology, are: Inductive learning, whole systems thinking, and individually optimized therapy. In honor of the ancient sages or "Rishis," whose voice is used to deliver Ayurvedic knowledge in the ancient texts of Ayurveda, these are referred to as the "Rishi principles." Individually optimized therapy is interpreted using the lens of decision analysis. Common research methodologies are examined for embodiment of these principles.

This paper aims at designing better performing feature-projection based classification algorithms and presents two new such algorithms. These algorithms are batch supervised learning algorithms and represent induced classification knowledge as feature intervals. In both algorithms, each feature participates in the classification by giving real-valued votes to classes. The prediction for an unseen example is the class receiving the highest vote. The first algorithm, OFP.MC, learns on each feature pairwise disjoint intervals which minimize feature classification error. The second algorithm. GFP.MC, constructs feature intervals by greedily improving the feature classification error. The new algorithms are empirically evaluated on twenty datasets from the UCI repository and are compared with the existing feature-projection based classification algorithms (FILIF, VFI5, CFP, k-NNFP, and NBC). The experiments demonstrate that the OFP.MC algorithm outperforms other feature-projection based classification algorithms. The GFP.MC algorithm is slightly inferior to the OFP.MC algorithm, but, if it is used for datasets with large number of instances, then it reduces the space requirement of the OFP.MC algorithm. The new algorithms are insensitive to boundary noise unlike the other feature-projection based classification algorithms considered here. (C) 2011 Elsevier Ltd. All rights reserved.

Growing amount of information in the world encourage the use of automatic data processing techniques that reduce humans routine work. There is a wide range of methods used for machine learning; however inductive learning algorithms are preferable in the systems where understanding of decision making steps and further processing of results is needed, for instance the expert systems, where the rules induced by learning algorithms can be used. As the classification tasks are getting more complicated computer program may not make enough informed decision by itself. In such situations collaborative approach between machine and systems user (expert) would be useful. Inductive learning system learns classification from training examples and uses induced rules for classifying new cases. If a decision cannot be inferred from rules base, a guess is performed. Interactive inductive system in uncertain conditions could ask human for decision and improve its knowledge base with the rule derived from this human-made decision. The paper summarises approaches discussed in related works and classifies them by the phase in inductive learning process in which the human interaction appears. As a result a new approach to interactive inductive system is presented. Conceptual example of topographical map classification using this system is demonstrated.

Cataloged from PDF version of article. This paper presents Feature Interval Learning algorithms (FIL) which represent multi-concept descriptions in the form of disjoint feature intervals. The FIL algorithms are batch supervised inductive learning algorithms and use feature projections of the training instances to represent induced classification knowledge. The concept description is learned separately for each feature and is in the form of a set of disjoint intervals. The class of an unseen instance is determined by the weighted-majority voting of the feature predictions. The basic FIL algorithm is enhanced with adaptive interval and feature weight schemes in order to handle noisy and irrelevant features. The algorithms are empirically evaluated on twelve data sets from the UCI repository and are compared with k-NN, k-NNFP, and NBC classification algorithms. The experiments demonstrate that the FIL algorithms are robust to irrelevant features and missing feature values, achieve accuracy comparable to the best of the existing algorithms with significantly less average running times. (C) 2010 Elsevier B.V. All rights reserved.

International audience; This paper presents an agent learning architecture based on the constructivist approach. After a brief introduction, we show details about the proposed model. Firstly, we explain how the agent represents the knowledge. We introduce an innovative learning method, inspired by the constructivist description of the human learning process. The kind of environment regularities that can be discovered by the method is analyzed. Finally, we discuss how an extension of this agent architecture can lead to the construction of abstract concepts.

International audience; Conditions complexity of management workshop of production manufacturing increases and requires heuristic algorithm adapted to the context. The principal difficulty then lies in the choice of heuristic algorithm to apply. We propose a help method to schedule, and more specifically a parameter setting up's help of an industrial scheduling software, being based on machine learning system able to extract knowledge from data. An inductive learning based on examples system is developed and replaced in a process of ECD (Extraction of Knowledge starting from Data). The first step pf this process is specific to our problem and uses in this case the capacities of simulation of a market software of scheduling.

This paper presents an application of machine learning algorithms based on inductive learning by logic minimization to the analysis of gene expression data. The characteristic properties of these data are a very large number of attributes (genes) and a relatively small number of examples (samples). Approaches to gene set reduction and to the detection of important disease markers are described. The results obtained on two well known publicly available gene expression classification problems are presented.

International audience; This work deals with fault detection and isolation (FDI) of an induction motor. Its supervision cannot be performed on the sole knowledge of analytical redundancy relations : a normal functioning state of the motor and a speed-sensor failure state cannot be distinguished from a behavioral analytical model. A solution is proposed using two inductive learning techniques based on decision tree formalism: C4.5 which is a milestone in top–down induction of decision trees, and BUST which is a solution for the functional separability problem of decision trees.

Genetic algorithms and evolution strategies are combined in order to build a multi-stage hybrid evolutionary algorithm for learning constrained approximate Mamdani-type knowledge bases from examples. The genetic algorithm niche concept is used in two of the three stages composing the learning process with the purpose of improving the accuracy of the designed fuzzy rule-based systems. The proposed genetic fuzzy rule-based system is used to solve an electrical engineering problem and the results obtained are compared with other methods presenting different characteristics.

International audience; Within the diagnosis assistance framework, the supervision of a system without behavioral analytical model requires a statistical model elaboratedfrom the observed data analysis. This approach is based on supervised learning techniques for large databases (data mining): from the knowledge of some parameters, the value of the variable to explain is predicted. The system dynamical behavior is reinjected in the initial database to be takeninto account by learning techniques which deal with raw data. C4.5, which represents the reference algorithm based on decision-tree formalism,is applied on a database from an induction motor in order to supervise it partially. More precisely, the problem consists in discriminating a normalfunctioning state of the motor from a speed sensor failure state.

This paper is a study of the problem of relevance in inductive concept learning. It gives definitions of irrelevant literals and irrelevant examples and presents efficient algorithms that enable their elimination. The proposed approach is directly applicable in propositional learning and in relation learning tasks that can be solved using a LINUS transformation approach. A simple inductive logic programming (ILP) problem is used to illustrate the approach to irrelevant literal and example elimination. Results of utility studies show the usefulness of literal reduction applied in LINUS and in the search of refinement graphs.

This paper presents a novel approach to the construction of reliable diagnostic rules from the available cases with known diagnoses. It proposes a simple and general framework based on the generation of the so-called confirmation rules. A property of a system of confirmation rules is that it allows for indecisive answers, which, as a consequence, enables that all decisive answers proposed by the system are reliable. Moreover, the consensus of two or more confirmation rules additionally increases the reliability of diagnostic answers. Experimental results in the problem of coronary artery disease diagnosis illustrate the approach.

Medical decision making based on inductive learning has been studied in order to collect experience necessary for practical use of such methods in clinical and epidemiological work. The decision trees have been constructed by using the modified Quinlan's approach based on choosing relevant attributes according to their informativity. An inductive learning software tool, ASSISTANT Professional, has been used for experimenting. The variability in results has been studied under varying learning conditions. Two sets of data have been chosen for learning experiments: from a study on rheumatoid factors in patients with rheumatoid arthritis, and from an epidemiological investigation of ageing. The results of this study indicate the necessity to determine inductive learning parameters for each particular problem. The pruning procedure is always recommended as it eliminates redundant elements in the tree. In problems with greater number of attributes, however, pruning itself is not guaranteeing satisfactory solutions. Interventions like the change of the minimal weight threshold might improve the situation. If these precautions are met, the method of inductive learning seems to be a useful guide in practical clinical and epidemiological decisions.

A system for inductive learning is used in automatic construction of decision trees. The results and experience in its use in the field of rheumatology are presented. The decision trees have been obtained under different conditions, i. e. without and with pruning. Variability of results has been noted and discussed. It is concluded that such decision trees could be a useful guide in clinical decisions.

Nowadays fuzzy logic controllers have been successfully applied to a wide range of engineering control processes. Several tasks have to be performed in order to design an intelligent control system of this kind for a concrete application. One of the most important and difficult ones is the extraction of the expert known knowledge of the controlled system. The aim of this paper is to present an evolutionary process based on genetic algorithms and evolution strategies for learning the fuzzy-logic-controller knowledge base from examples in three different stages. The process allows us to generate two different kinds of knowledge bases, descriptive and approximate ones, depending on the scope of the fuzzy sets giving meaning to the fuzzy-control-rule linguistic terms, taking preliminary linguistic-variable fuzzy partitions as a base. The performance of the method proposed is shown by measuring the accuracy of the fuzzy logic controllers designed in the fuzzy modeling of three three-dimensional surfaces presenting different characteristics, and by comparing them with others generated by means of three methods based on Wang and Mendel's knowledge-base generation process.

The rapid development of new learning algorithms increases the need for improved accuracy estimation methods. Moreover, methods allowing the comparison of several different learning algorithms are important for the performance evaluation of new ones. In this paper we propose new accuracy estimation methods which are extensions of the k-fold cross-validation method. The methods proposed construct cross-validation folds deterministically instead of using the random sampling approach. The deterministic construction of folds is performed using unsupervised stratification by exploiting the distribution of instances in the instance space. Our methods are based either on the one-center approach or on clustering procedures. These methods attempt to construct more representative folds, therefore reducing the bias of the resulting estimator. At the same time, our methods allow direct comparisons between the performance of learning algorithms in different experiments, since no randomness is present. A simulation experiment examining the performance of the proposed methods is reported, depicting their behavior in a variety of situations. The new methods reduce mainly the bias of the estimator. © 2000 Elsevier Science B.V. All rights reserved.

In this paper, a novel algorithm for learning structured descriptions, ascribable to the category of symbolic techniques, is proposed. It faces the problem directly in the space of the graphs, by defining the proper inference operators, as graph generalization and graph specialization, and obtains general and coherent prototypes with a low computational cost with respect to other symbolic learning systems. The proposed algorithm is tested with reference to a problem of handwritten character recognition from a standard database.