Showing papers on "Active learning (machine learning) published in 1989"

Generic constraints on underspecified target trajectories

Abstract: Although general network learning rules are of undeniable interest, it is generally agreed that successful accounts of learning must incorporate domain-specific, a priori knowledge. Such knowledge might be used, for example, to determine the structure of a network or its initial weights. The author discusses a third possibility in which domain-specific knowledge is incorporated directly in a network learning rule via a set of constraints on activations. The approach uses the notion of a forward model to give constraints a domain-specific interpretation. This approach is demonstrated with several examples from the domain of motor learning. >

An experimental comparison of symbolic and connectionist learning algorithms

Abstract: Despite the fact that many symbolic and connectionist (neural net) learning algorithms are addressing the same problem of learning from classified examples, very little Is known regarding their comparative strengths and weaknesses This paper presents the results of experiments comparing the ID3 symbolic learning algorithm with the perceptron and back-propagation connectionist learning algorithms on several large real-world data sets The results show that ID3 and perceptron run significantly faster than does backpropagation, both during learning and during classification of novel examples However, the probability of correctly classifying new examples is about the same for the three systems On noisy data sets there is some indication that backpropagation classifies more accurately

Noise-tolerant instance-based learning algorithms

Abstract: Several published reports show that instancebased learning algorithms yield high classification accuracies and have low storage requirements during supervised learning applications. However, these learning algorithms are highly sensitive to noisy training instances. This paper describes a simple extension of instancebased learning algorithms for detecting and removing noisy instances from concept descriptions. This extension requires evidence that saved instances be significantly good classifiers before it allows them to be used for subsequent classification tasks. We show that this extension's performance degrades more slowly in the presence of noise, improves classification accuracies, and further reduces storage requirements in several artificial and real-world database applications.

Experimental comparison of human and machine learning formalisms

Abstract: In this paper we describe the results of a set of experiments in which we compared the learning performance of human and machine learning agents. The problem involved the learning of a concept description for deciding on the legality of positions within the chess endgame King and Rook against King. Various amounts of background knowledge were made available to each learning agent. We concluded that the ability to produce high performance in this domain was almost entirely dependent on the ability to express first-order predicate relationships.

Rule creation and rule learning through environmental exploration

Abstract: The task of learning from environment is specified. It requires the learner to infer the laws of the environment in terms of its percepts and actions, and use the laws to solve problems. Based on research on problem space creation and discrimination learning, this paper reports an approach in which exploration, rule creation and rule learning are coordinated in a single framework. With this approach, the system LIVE creates STRIPS-Iike rules by noticing the changes in the environment when actions are taken, and later refines the rules by explaining the failures of their predictions. Unlike many other learning systems, since LIVE treats learning and problem solving as interleaved activities, no training instance nor any concept hierarchy is necessary to start learning. Furthermore, the approach is capable of discovering hidden features from the environment when normal discrimination process fails to make any progress.

A system for incremental learning based on algorithmic probability

Abstract: We have employed Algorithmic Probability Theory to construct a system for machine learning of great power and generality. The principal thrust of present research is the design of sequences of problems to train this system. Current programs for machine learning are limited in the kinds of concepts accessible to them, the kinds of problems they can learn to solve, and in the e‐ciency with which they learn | both in computation time needed and/or in amount of data needed for learning. Algorithmic Probability Theory provides a general model of the learning process that enables us to understand and surpass many of these limitations. Starting with a machine containing a small set of concepts, we use a carefully designed sequence of problems of increasing di‐culty to bring the machine to a high level of problem solving skill. The use of training sequences of problems for machine knowledge acquisition promises to yield Expert Systems that will be easier to train and free of the brittleness that characterizes the narrow specialization of present day systems of this sort. It is also expected that the present research will give needed insight in the design of training sequences for human learning.

Learning control with neural networks

Abstract: A neural control model based on learning of the system inverse is proposed. Learning control is a control method wherein experience gained from previous performance is automatically used to improve future performance. A learning scheme called the inverse transfer learning scheme is introduced. Compared to previous learning schemes, this scheme provides faster convergences to the minimum error state and reflects properties of highly coupled nonlinear dynamic systems. The scheme is applied to the pole-balancing control problem through computer simulation to demonstrate control capability. >

Knowledge Representation and Organization in Machine Learning

Abstract: Explanation: A source of guidance for knowledge representation.- (Re)presentation issues in second generation expert systems.- Some aspects of learning and reorganization in an analogical representation.- A knowledge-intensive learning system for document retrieval.- Constructing expert systems as building mental models or toward a cognitive ontology for expert systems.- Sloppy modeling.- The central role of explanations in disciple.- An inference engine for representing multiple theories.- The acquisition of model-knowledge for a model-driven machine learning approach.- Using attribute dependencies for rule learning.- Learning disjunctive concepts.- The use of analogy in incremental SBL.- Knowledge base refinement using apprenticeship learning techniques.- Creating high level knowledge structures from simple elements.- Demand-driven concept formation.

A structural learning algorithm with forgetting of link weights

Abstract: Summary form only given, as follows. Backpropagation learning suffers from serious drawbacks: first the necessity of a priori specification of a model structure, and second the difficulty in interpreting hidden units. To cope with these drawbacks the author proposes a novel learning algorithm, called structural learning algorithm, which generates a skeletal structure of a network: a network in which minimum number of links and a minimum number of hidden units are actually used. The resulting skeletal structure solves the first difficulty of trial and error. It also solves the second difficulty due to its clarity. In addition to these two benefits, the structural learning algorithm is also advantageous in dealing with a network composed of multiple modules. It explains how links from other modules emerge, while pruning those from the outside world full of redundant information. >

Cost-sensitive concept learning of sensor use in approach and recognition

Abstract: Publisher Summary This chapter explores a prototype learning method that complements recent work in incremental learning by considering the role of external costs arising from realistic environmental assumptions. It would be logical to use machine learning techniques to develop a system that builds an efficient recognition process in response to available sensors and the perceptual qualities of objects encountered when aiming for a reduction of laborious encoding process. This learning-from-examples task is similar to others that machine learning has addressed save for two salient differences, both of which hinge on the cost aspects of sensing and acting. Observations are virtually very large—each object may be described by the results of the many instantiations of all possible sensing procedures. As sensing procedures have different costs, that is, different execution times, the ability of a particular sensor feature to discriminate between appropriate actions must be balanced against the cost of its corresponding sensing procedure.

Learning classification rules using Bayes

Abstract: This paper considers the problem of learning classification rules from data in the context of knowledge acquisition. Bayesian theory provides a framework for both designing learning algorithms and for approaching specific learning applications, for instance, in the selection and tuning of learning tools. Experiments are reported demonstrating how this can be done using two well known learning approaches: simple Bayes classifiers and decision trees.

Combining empirical and analytical learning with version spaces

Abstract: This paper describes an approach to combining empirical and analytical learning using incremental version-space merging (Hirsh, 1989). The basic idea is to use analytical learning to generalize training data before doing empirical learning. The combination operates like empirical learning given no knowledge, but can utilize knowledge when provided, and thus exhibits behavior along a spectrum from knowledge-poor to knowledge-rich learning.

On learning from exercises

Abstract: This paper explores a new direction in the formal theory of learning - learning in the sense of improving computational efficiency as opposed to concept learning in the sense of Valiant. Specifically, the paper concerns algorithms that learn to solve problems from sample instances of the problems. We develop a general framework for such learning and study the framework over two distinct random sources of sample instances. The first source provides sample instances together with their solutions, while the second source provides unsolved instances or “exercises”. We prove two theorems identifying conditions sufficient for learning over the two sources, our proofs being constructive in that they exhibit learning algorithms. To illustrate the scope of our results, we discuss their application to a program that learns to solve restricted classes of symbolic integrals.

Iterative learning for trajectory control

Abstract: Learning control is an iterative approach to the problem of improving transient behavior for processes that are repetitive in nature. A complete analysis of the learning control problem is given for the case of linear, time-invariant plants and controllers. The analysis offers insights into the nature of the solution of learning control schemes. First, an approach based on parameter estimation is given. Then, it is shown that for finite-horizon problems it is possible to design a learning control algorithm which converges in one step. A brief simulation example is presented to illustrate the effectiveness of iterative learning for controlling the trajectory of a nonlinear robot manipulator. >

Learning control in robotic systems

Abstract: The concept of learning and training machines, and some early methodologies were introduced about two decades ago. Robotic systems, undoubtedly, can be developed to a more advanced and intelligent stage. The realization of the learning capability, analogous to the human learning and thinking process, is a desired primary function. A betterment process has been investigated in the literature for applications of learning control to robotic systems. The existing schemes are a type of one-step learning process. A neighboring (2m+1)-step learning control scheme for robotic systems is presented in this paper. For each process, a betterment algorithm which chooses a generalized momentum as an output function, is executed. Also, it is associated with a conceptual learning process by adding a self-teaching knowledge base for speeding up the convergence, so that the learning capability of the resulting robotic systems can be enhanced.

A frequency-domain approach to learning control: implementation for a robot manipulator

Abstract: A frequency-domain approach to the analysis and design of a learning control law for linear dynamic systems is presented. In its most simple version the scheme uses two separate filters in order to achieve rapid improvements in a specified bandwidth while cutting off-perhaps unmodeled-dynamic effects which would bar the convergence. The merit of this approach is to make explicit the tradeoff between global convergence conditions and approximate learning of trajectories. The proposed learning controller can also be applied to robot manipulators for exact tracking of repetitive trajectories. The scheme has been implemented successfully for learning the actuator inputs which permit accurate reproduction of robot trajectories defined in the joint space. Experimental results are reported which also show the efficacy of this learning scheme in avoiding the possible occurrence of an unstable behavior. >

CALM networks: a modular approach to supervised and unsupervised learning

Abstract: The authors discuss some problems in learning networks. They propose a new learning procedure, CALM (Categorizing And Learning Module). CALM uses pairs of excitatory representation nodes and inhibitory veto nodes, bound together in a modular structure with an arousal node. Learning in the module is enhanced by a nonspecific external node connected to the arousal node. The system is capable of both supervised and unsupervised learning and can both discriminate and generalize across similar patterns. A system constructed out of several CALM modules is shown to learn the XOR relationship with supervised and unsupervised presentation. It also models list recall and word completion memory tasks and can learn, unsupervised, handwritten digits and recognize them with unknown authors. >

Inductive inference with bounded number of mind changes

Abstract: Inductive inference machines (IIMs) synthesize programs, given their intended input-output behavior. The program synthesis is viewed as a potentially infinite process of learning by example. Smith [20] studied team learning and obtained results that characterized trade-offs between the number of machines and resources in the learning process. Pitt [16] defined probabilistic learning and showed that ‘probabilistic learning’ is the same as ‘team learning’. Later [17] introduced probabilistic team learning and compared probabilistic team learning and team learning. However, for any given team when we restrict amount of resources allotted for each IIM, then most of the above results fail to hold. This paper studies the relationships between team learning, probabilistic learning and probabilistic team learning when limited resources are available. Some preliminary results obtained indicates a very interesting relationship between them. The proofs for some of the preliminary results, used n-ary recursion theorems, and some complex diagonalization.

The light bulb problem

Abstract: In this paper, we consider the problem of correlational learning and present efficient algorithms to determine correlated objects.

Neural-style microsystems that learn

Abstract: The basic operation of biological and electronic (artificial) neural networks (NNs) is examined. Learning by NNs is discussed, covering supervised learning, particularly back-propagation, and unsupervised and reinforcement learning. The use of VLSI implementation to speed learning is considered briefly. Applications of neural-style learning chips to pattern recognition, data compression, optimization, and expert systems is discussed. Problem areas and issues for further research are addressed. >

Using a context-sensitive learning for robot arm control

Abstract: A class of networks called context-sensitive learning networks is proposed for use in the learning of complex nonlinear mappings. Particular attention is given to a network architecture for learning to control a robot arm by learning independently the different entries of the inverse Jacobian matrix. Computer simulation results show that the network is able to learn the inverse Jacobian of the PUMA 560 arm for inverse kinematic control. The network also generalized well when unseen testing examples are presented to it. >

Design techniques for transient response control

Abstract: This dissertation analyzes and develops some design techniques for the control of transient response characteristics. First, a method for minimizing the output overshoot in SISO discrete-time systems with fixed-order controllers is given. The technique finds the optimal location of the assignable closed-loop system zeros for a given set of poles. Second, we investigate the use of multirate sampling techniques for assigning all the poles and zeros of the closed-loop system. Results are given for SISO plants, for square MIMO plants, and for the SISO servo problem. Intersample effects are also analyzed. Next, we consider a new approach called learning control. Learning control is an iterative approach to deriving the output of an optimal pre-filter, using actual input/output data. A complete analysis of the learning control problem is given for LTI systems, including a study of the nature of the solution, development of a learning controller based on parameter estimation, and a study of finite-horizon learning control schemes. We also present a learning control scheme that can be applied to a class of nonlinear systems which includes some models of robotic manipulators. Finally, we discuss the potential applications of artificial neural networks to the learning control problem.

Input space representation for refinement learning control

Abstract: A learning control approach called refinement, in which a fixed controller is first designed using analytic design tools is explored. This controller's performance is refined by a secondary learning controller, which is a reinforcement learning-based connectionist network. The issue is the representation of the input space of the refinement learning controller. In previous work, the input space was quantized into fixed boxes and each box became a control situation for the learning controller. The drawback was that the learning control designer had to know how to quantize the space. An approach to eliminating the quantization of the input space is described. The new input space representation consists of functions that act as receptive fields and have the shape of multivariate Gaussian probability density functions; they are the first layer in the learning network. Experiments used a tracking control problem with an additive nonlinearity. The learning controller adds an appropriate control signal on the basis of a given evaluation function, in order to improve the fixed controller's ability to track a reference signal. >

When Will Machines Learn

Abstract: Why don’t our learning programs just keep on going and become generally intelligent? The source of the problem is that most of our learning occurs at the fringe of what we already know. The more you know, the more (and faster) you can learn.

Discovering problem solving strategies: what humans do and machines don't (yet)

Abstract: An unusually detailed analysis and simulation of a human problem solving protocol uncovered 10 cases of strategies being discovered. Although most of these learning events were adequately modeled by existing machine learning techniques, several present interesting challenges for machine learning research. This paper briefly presents the experiment and the 10 learning events. The protocol analysis is detailed in VanLehn (1989). The simulation system, TETON, is described in VanLehn and Ball (in press).

The acquisition of model-knowledge for a model-driven machine learning approach

Abstract: Knowledge acquisition systems with a model-driven learning mechanism require the representation of that model in the system. The model which guides the learning mechanism must be distinguished from the knowledge (domain model) which is to be learned with the learning mechanism; only the former is the concern of this paper. If the model for guiding the learning mechanism is to be enlarged and improved while working with such a system, the acquisition and representation of new parts of this model must be supported. In addition to the insertion of new parts into the existing model, it is very important to consider redundancy, integrity and completion, because the quality of the model influences the quality of the learning capabilities of the knowledge acquisition system.

A machine learning approach to student modelling

Abstract: This paper describes an application of established machine learning principles to student modelling. Unlike previous machine learning based approaches to student modelling, the new approach is based on attributevalue machine learning. In contrast to many previous approaches it is not necessary for the lesson author to identify all forms of error that may be detected. Rather, the lesson author need only identify the relevant attributes both of the tasks to be performed by the student and of the student’s actions. The values of these attributes are automatically processed by the student modeler to produce the student model.

Learning Disjunctive Concepts

Abstract: We have presented in this paper a learning system that is able to produce characteristic and discriminant versions of disjunctive target concepts. The algorithm is able to process all classes at once. It integrates several inductive learning methods: An ID3-based system (with several important improvements over the original ID3) and a generalization engine. The first way we approached integration was to use a switchbox mecanism that enabled us to quickly identify strong and weak points of each learning systems. The switchbox integration was also used to test plausible ways of fully merging the systems. When the limit of what could possibly be done was reached, we then went on integrating totally the learning systems. This required major changes in the systems themselves and was achieved by designing a powerful object-based representation formalism and implementing a common pattern-matcher. Following this work, several problems remained to be solved and lead us to experiment with other learning systems such as the star algorithm. The integrated system can use domain knowledge and is being applied on large scale problems.