T.G. Dietterich, B. London, K. Clarkson & G. Dromey, `Learning and inductive inference', Handbook of Artificial Intelligence, Vol. III, P. Cohen & E.A. Feigenbaum (eds.), William Kaufmann, 1982.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....who find themselves having to estimate complexity must resort to informal and heuristic methods of evaluation. A longstanding complexity estimation heuristic is simply to determine whether or not the learning problem in question involves the capturing of relations. As has been known for some time [16] problems which involve relational effects are, in general, harder to solve than problems which do not. Another, reliable heuristic approach involves determining the order of statistical effect which underpins the solution to the problem [17] The idea here is that higher order effects are harder ....

Dietterich, T., London, B., Clarkson, K. and Dromey, G. (1982). Learning and inductive inference. In P. Cohen and E. Feigenbaum (Eds.), The Handbook of Artificial Intelligence: Vol III. Los Altos: Kaufmann.

....an S set the set of most specific concept definitions that cover at least one of the instances, and as a G set the universal concept that includes everything. If the singlerepresentation trick holds (i.e. for each instance there is a concept definition whose extension only contains that instance) [Dietterich et al. 1982], the S set contains only the instance and its neighboring instances. If an instance is negative the S set contains the empty concept that includes nothing and the G set contains all minimal specializations of the universal concept that excludes the instance or one of its neighbors. 4.1 Searching ....

T. G. Dietterich, B. London, K. Clarkson, and G. Dromey. Learning and inductive inference. In P. Cohen and E. A. Feigenbaum, editors, The Handbook of Artificial Intelligence, Volume III. William Kaufmann, Los Altos, CA, 1982.

....hand. What language should be used to describe an individual We employ the commonly used single representation trick wherein the description of an individual is itself a Classic sentence. This approach is supported by the description logic community (Borgida, 1992; Bobrow Winograd, 1977; Dietterich et al. 1982), in which it is often convenient and desirable to represent concepts and examples using the same language. In fact, Cohen and Hirsh (1994a) note that in many implemented description logics, it is possible to attach an arbitrary description to an instance [example] hence the distinction between ....

....Classic descriptions is lost. If we are given a target Classic description, which other Classic descriptions then are positive examples, and which are negative examples Again, following work in the description logic community (Cohen Hirsh, 1994a; Borgida, 1992; Bobrow Winograd, 1977; Dietterich et al. 1982), we define a positive example to be any Classic description that denotes, for every possible interpretation, a subset of those individuals denoted by the target description. Thus, each positive example has denotation that is a subset of the denotation of the target concept, regardless of the ....

Dietterich, T. G., London, B., Clarkson, K., & Dromey, G. (1982). Learning and inductive inference. In The Handbook of Artificial Intelligence, Volume 3. William Kaufmann.

....drop. 3. New information interferes with control knowledge implicitly encoded in some structure imposed on the knowledge base, such as the order in which rules appear. For example, it has been noted that simply adding new rules to MYCIN caused existing rules to apply erroneously or not at all [Die82, page 331] 1 These conflicts illustrate one general category of interaction that occurs between new and prior knowledge. New information can also interact synergistically with existing knowledge. Each synergistic interaction suggests new knowledge not explicitly contained in the new ....

....prescribe admissibility criteria, see Section B. 1) The application of these postulates is illustrated with the multipleextension problem [Gin87] one of the essential issues in the study of the nonmonotonic nature of knowledge (and related to the credit assignment problem in machine learning [Die82] Consider the following learning situation: prior knowledge: 1) a ) b (2) b ) c (3) c ) d (4) a new information: d Which prior belief should be retracted Any three out of the four initial beliefs are consistent with the new information, so there are many possible extensions; however, ....

[Article contains additional citation context not shown here]

T.G. Dietterich. Learning and inductive inference. In P.R. Cohen and E.A. Feigenbaum, editors, The Handbook of Artificial Intelligence, volume Volume 3. Los Altos, CA: William Kaufmann, Inc., 1982.

....learning component, and again for the other components of the system. One solution to this problem is to enable the machine learning program to choose its own vocabulary. This approach is usually called constructive induction [Michalski, 1983] although it is also called the new term problem [Dietterich, London, Clarkson Dromey, 1982], feature extraction [Kittler, 1986] and feature generation [Rendell, 1985] Most implementations are incremental; whenever the program is not making reasonable progress in achieving some goal, domain independent heuristics are used to change the vocabulary. Constructive induction may be faster ....

Dietterich, T. G., London, B., Clarkson, K., & Dromey, G. (1982). Learning and inductive inference. In Cohen & Feigenbaum (Eds.), The Handbook of Artificial Intelligence: Volume III. San Mateo, CA: Morgan Kaufmann.

....learning rule is guaranteed to find a hyperplane that separates the positive examples from the negative examples, yielding 100 classification accuracy. When the examples are not linearly separable, the least mean squares learning rule tends to find hyperplanes that minimize the mean squared error [Dietterich, London, Clarkson Dromey, 1982]. These hyperplanes provide good, but not perfect, classification. In each experiment, training and testing were alternated until the LTU reached a specified level of accuracy in classifying examples. Each training phase was performed on only a very small set of randomly chosen examples, but ....

Dietterich, T. G., London, B., Clarkson, K., & Dromey, G. (1982). Learning and inductive inference. In Cohen & Feigenbaum (Eds.), The Handbook of Artificial Intelligence: Volume III. San Mateo, CA: Morgan Kaufmann.

....examples with techniques for construction of new features, in order to solve difficult problems in learning from examples. 1. Introduction. The problem of new terms, also known as the constructive induction problem, has long been considered a source of difficulty in machine learning (Dietterich, 1982). Simple classifiers using only the primitive features of description have limited learning capabilities. For example: i) Single layered neural networks can realize only those class dichotomies, where the classes are linearly separable in the feature space (Minsky, 1969) ii) Selective induction ....

....(Rumelhart, 1986) can implement learning in multi layered networks. They implicitly create weighted combinations of primitive features in the internal (hidden) units of such networks. Constructive induction, on the other hand, explicitly constructs and tests new terms from the primitive features (Dietterich et al. 1982) by applying feature construction operators. Both of these approaches transform the primitive feature space of the problem into another in which the classes to be discriminated are separable using simple discrimination surfaces. With a few exceptions (Muggleton, 1988) these techniques provide no ....

T. G. Dietterich, B. London, K. Clarkson, and G. Dromey, "Learning and Inductive Inference," in The Handbook of Artificial Intelligence, ed., P. R. Cohen and E. A. Feigenbaum. Kaufmann, 1982.

.... for a statistical effect in the form of an observed probability) If it was not obvious before, the introduction of this new terminology should drive home the point that this analysis gives a foundation to the well known Machine Learning heuristic which states that learning relationships is hard [4]. 5 Example To illustrate this notion of indirect justification, consider the following training set. This is based on two input variables (x1 and x2) and one output variable (y1) There are six training examples in all. An arrow separates the input part of the example from the output part. x1 ....

Dietterich, T., London, B., Clarkson, K. and Dromey, G. (1982). Learning and inductive inference. In P. Cohen and E. Feigenbaum (Eds.), The Handbook of Artificial Intelligence: Vol III. Los Altos: Kaufmann.

....of inferring general hypotheses from specific information. It is the main inference step in concept learning from examples, which is the problem of finding a concept definition that correctly classifies all positive and negative examples, while maximising expected accuracy on unseen instances [4]. Induction can also be employed to infer non classificatory hypotheses, that describe properties (rather than definitions) of the data. From the database perspective one would refer to such properties as integrity constraints. Since attribute dependencies constitute a form of database ....

....induction algorithms construct hypotheses directly from the data. They are less easily viewed as search algorithms, and are in some cases even deterministic. For instance, when inducing a definition for the append predicate we may encounter the facts append( a] a] and append( 1,2] [3,4], 1,2,3,4] If we assume that these are generated by the same clause, we can construct the clause head append( A B] C, A D] by anti unification, and then proceed to find appropriate restrictions for the variables. Since the data can be viewed as extremely specific hypotheses, such algorithms ....

[Article contains additional citation context not shown here]

T.G. Dietterich, B. London, K. Clarkson & G. Dromey, `Learning and inductive inference', Handbook of Artificial Intelligence, Vol. III, P. Cohen & E.A. Feigenbaum (eds.), William Kaufmann, 1982.

....new features, then runs the algorithm again. The human is therefore part of the learning cycle, so the process is not automatic. The problem of devising new terms for an induction task has been given many names. It has been called constructive induction [Michalski, 1983] and the new term problem [Dietterich, London, Clarkson Dromey, 1982], to emphasize the fact that the induced concept is formed from newly constructed terms, rather than from those that were initially given to it. The set of representable concepts has been called a bias of the learning method, so the process of adding new terms has been called shift of bias [Utgoff ....

Dietterich, T. G., London, B., Clarkson, K., & Dromey, G. (1982). Learning and inductive inference. In Cohen & Feigenbaum (Eds.), The Handbook of Artificial Intelligence: Volume III. San Mateo, CA: Morgan Kaufmann.

....do not appear in this paper, but are to be found instead in [ Grosof, 1993 ] That paper thus acts as a companion to this paper. An important common sense capability is to learn by taking advice , or being told , as proposed, for example, by [ McCarthy, 1968 ] and discussed, for example, in [ Dietterich, 1982 ] By taking advice, we mean adopting things one is told as one s own beliefs, more or less directly. Advice may arise from a variety of sources of information: messages from various other agents (e.g. an inputting user ) reading various texts, etc. Such learned or assimilated 1 knowledge ....

Thomas G. Dietterich. Learning and inductive inference. In Paul R. Cohen and Edward A. Feigenbaum, editors, The Handbook of Artificial Intelligence, Volume 3, pages 323--512. Morgan Kaufmann, San Mateo, California, 1982.

....a machine s ability to reason is tied directly to its representation of a problem. Since the beginning of the study of Artificial Intelligence at the Dartmouth Conference in 1956, researchers have studied how a machine can learn, and many successful learning algorithms have been developed (see Dietterich, London, Clarkson and Dromey (1982) for a history of the field) A problem that has received less attention is how to select which learning algorithm to use for a given task. Typically, when given a learning task, a human selects an input representation for the examples and a learning algorithm. The ability of the chosen algorithm ....

Dietterich, T. G., London, B., Clarkson, K., & Dromey, G. (1982). Learning and inductive inference. In Cohen & Feigenbaum (Eds.), The Handbook of Artificial Intelligence: Volume III. San Mateo, CA: Morgan Kaufmann.

....this proposal examines function approximation with an emphasis on issues faced in an optimal control framework adopted for the research. Although forming abstractions is consistent with a common theme in learning, namely, transforming data representations to improve subsequent processing (Dietterich, 1982; Hinton, 1989) it appears that most learning research on function approximation methods fails to address the full range of issues present in optimal control. Learning from examples is undeniably important, but if agents are to learn autonomously to act in dynamic environments, then optimal ....

Dietterich, T. G. (1982). Learning and inductive inference. In Cohen, P. R. & Feigenbaum, E. A. (Eds.), The Handbook of Artificial Intelligence, chapter XIV, pages 323--511. Redwood City, CA: Addison-Wesley.

....[25] In all of the above problems, the system learns a general class description from instances of the class, and therefore, this type of inductive learning can be called instance to class generalization. A review of several methods for such instance to class generalization can be found in [8, 24]. Another type of inductive learning involves constructing a description of a whole object by observing only selected parts of it. For example, given a set of fragments of a scene, the problem is to hypothesize the description of the whole scene. A very important case of such pan to whole ....

....however, may not know the proper transformations a priori. These learning systems will need to select or invent appropriate task oriented transformations for each learning situation. This description space transformation problem has been called by various authors the data interpretation problem [8] or the reformulation problem [2] We prefer the term task oriented transformation problem, since it emphasizes that the proper choice of data transformations depends upon the task being performed. In the prediction problem discussed here, the desired sequencegenerating rules are described in a ....

Dietterich, T.G., London, R., Clarkson, K. and Dromey, G., Learning and inductive inference, in: P.R. Cohen and E.A. Feigenbaum (Eds.), The Handbook o/ Artificial Intelligence 3 (Kaufmann, Los Altos, CA, 1982) Ch. 14.

No context found.

T.G. Dietterich, B. London, K. Clarkson & G. Dromey, `Learning and inductive inference', Handbook of Artificial Intelligence, Vol. III, P. Cohen & E.A. Feigenbaum (eds.), William Kaufmann, 1982.

No context found.

Dietterich, T., London, B., Clarkson, K. and Dromey, G. (1982). Learning and inductive inference. In P. Cohen and E. Feigenbaum (Eds.), The Handbook of Artificial Intelligence: Vol III. Los Altos: Kaufmann.

No context found.

Dietterich, T., London, B., Clarkson, K. and Dromey, G. (1982). Learning and inductive inference. In P. Cohen and E. Feigenbaum (Eds.), The Handbook of Artificial Intelligence: Vol III. Los Altos: Kaufmann.

No context found.

Dietterich, T., London, B., Clarkson, K. and Dromey, G. (1982). Learning and inductive inference. In P. Cohen and E. Feigenbaum (Eds.), The Handbook of Artificial Intelligence: Vol III. Los Altos: Kaufmann.

No context found.

T. G. Dietterich, B. London, K. Clarkson, and G. Dromey. Learning and inductive inference. In P. Cohen and E. A. Feigenbaum, editors, The Handbook of Artificial Intelligence, Volume III. William Kaufmann, Los Altos, CA, 1982.

No context found.

Dietterich, T., London, B., Clarkson, K. and Dromey, G. (1982). Learning and inductive inference. In P. Cohen and E. Feigenbaum (Eds.), The Handbook of Artificial Intelligence: Vol III. Los Altos: Kaufmann.

No context found.

Dietterich, T., London, B., Clarkson, K. and Dromey, G. (1982). Learning and inductive inference. In P. Cohen and E. Feigenbaum (Eds.), The Handbook of Artificial Intelligence: Vol III. Los Altos: Kaufmann.

No context found.

Dietterich, T., London, B., Clarkson, K. and Dromey, G. (1982). Learning and inductive inference. In P. Cohen and E. Feigenbaum (Eds.), The Handbook of Artificial Intelligence: Vol III. Los Altos: Kaufmann.

No context found.

Dietterich, T., London, B., Clarkson, K. and Dromey, G. (1982). Learning and inductive inference. In P. Cohen and E. Feigenbaum (Eds.), The Handbook of Artificial Intelligence: Vol III. Los Altos: Kaufmann.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC