Tom Mitchell. Generalization as Search. Artificial Intelligence, 18(2):203--226, 1982.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....a hypothesis space useful for learning (where the hypothesis space of a learner refers to the set of possible hypotheses that it can generate. 2. 4 Desirable properties for a hypothesis space Learning can be thought of as search through the hypothesis space for a hypothesis that fits the data [ Mitchell, 1981 ] In principle, weak methods like enumeration can be used to perform this search; however, most practical learning systems make use of special properties of the hypothesis space in order to search the space more efficiently. The following have empirically proven to be desirable properties for a ....

Tom Mitchell. Generalization as search. In Readings in Artificial Intelligence. Morgan Kaufmann, 1981.

....of machine learning. In particular, constraint acquisition is regarded as search through an hypothesis space of constraints over which a general to speci c ordering is explicitly known, or is implicit in its representation. Examples provided by the user can be used to identify a version space ([4]) of constraints that the user could be attempting to articulate. We employ an interactive model similar to that proposed by Sammut and Banerji for acquiring constraints from the user [9] As the user provides examples of what he would nd acceptable (positive examples) we attempt to generalize ....

....examples are classi ed as negative, positive examples are classi ed as positive. The term version space is used to describe the subset of the hypothesis space consistent with the training examples. Readers who are unfamiliar with version space learning are encouraged to refer to the literature [4, 5]. In Section 2.1 we present the model of interaction used in this work. An example of an interactive session involving the acquisition of a simple placement constraint is presented in Section 2.2. 2.1 A Model for Interactive Constraint Acquisition The model of interaction employed here is as ....

Tom Mitchell. Generalization as search. Arti cial Intelligence, 18(2):203-226, 1982.

....is introduced with FOIL in [Quib] FOIL is a top down system, it starts with the most general rules and learns by specializing. Bottom up logic programming is introduced with GOLEM in [MF92] Learning in GOLEM proceeds by generalizing. Version Spaces as a learning algorithm is introduced in [Mit82]. Version Spaces are studied further in [Hir91] Certainty based active learning is described in [CAL94] and [CGJ96] Another kind of active learning, which is committee based, in [DE95] and [LT97] Active learning is widely used in conjunction with other machine learning techniques. Active ....

Tom Mitchell. Generalization as search. Artificial Intelligence, 1982.

....without explaining all of the behavior in each observation. In addition, the representation itself can be used to obtain a measure of the diversity of behavior encapsulated in the observed agent actions. Conceptually, Behavioral Bounding is very similar to Mitchell s Version Space framework [4]. Given a language to describe an agent s behavior, and a general to speci c ordering over representations in this language, we can construct a maximally speci c representation of the observed agent behavior. Figure 2 illustrates this concept. Each Most Specific Common Generalization Behavior ....

Tom M. Mitchell. Generalization as search. Arti cial Intelligence, 18(2):203-226, 1982.

....types of learning. In supervised learning, an agent observes input output pairs as examples, and learns about the function mapping the inputs to the outputs. The values of the outputs are provided by the environment, which can be thought of as a supervisor or a teacher. Decision tree methods [15, 11] and neural network learning [6] fall into this category. In unsupervised learning, the agent is given a collection of input values but no output values. The agent has to find regularity in the inputs by itself. Clustering [5] and discovery [8] methods fall into this category. Unsupervised ....

Tom Mitchell. Generalization as search. Artificial Intelligence, 18(2):203--226, 1982.

.... rely on a certain notion of simplicity (cf Aristotles, Occam s razor (William of Ockham [1285 1347] Entia non sunt multiplicandi sine necessitae Entities should not be multiplied beyond necessity In the context of Machine Learning, the way out if this paradox is the use of a learning bias [Mit82]. 1.1.2 Towards a de nition of simplicity 1.1.3 The Learning dilemma 1.1.4 Solomono s approach 1.2 Algorithmic Information Theory The notion of algorithmic complexity (or algorithmic information) has been independently introduced by three authors : Ray Solomono was looking for a way to ....

Tom M. Mitchell, Generalization as search, Articial Intelligence 18 (1982), 203-226.

....proach uses exploration and experimentation to fill in missing pre conditions and e#ects. Wang s OBSERVER system [12] is able to work from an empty knowledge base, using traces from experts problem solving to fill in the empty operator descriptions using an approach similar to the version method [9]. Mahadevan showed how learning action models can enable transfer in reinforcement learning [8] One of the limitations of these approaches is that they assume pre conditions are conjunctive and they do not handle domain noise. TRAIL [3] extends this work by allowing disjunctions in the ....

Tom M. Mitchell. Generalization as search. Artificial Intelligence, 18:203--226, 1982.

....Natarajan [21] in terms of one sided learning. As another example, we could consider the class of learners that only select a minimal consistent hypothesis. This corresponds to requiring that the learner always selects an element from the set of most specific concepts in Mitchell s version space [20]. 30 Another interesting variation is one in which the learner is not required to exactly identify the target, but rather needs only output an ffl good approximation to the target. Romanik and Smith [23, 24] consider a PAC style criterion in their work. Finally, another interesting model is ....

Tom M. Mitchell. Generalization as search. Artificial Intelligence, 18(2):203--226, 1982.

....symbol. The goal is to predict what symbol is correct for a given set of predicates. Many approaches have been taken to solve this problem. The algorithm ID3 [36] builds a decision tree using the predicates as choice points. Another solution to the problem was Mitchell s concept space algorithm [25]. A concept space represents the set of all solutions as a graph and each positive and negative example cuts off portions of the graph until one is left with a single concept in the space. Classifier algorithms have tended to use only examples devoid of annotation. The reason for this is that ....

Tom M. Mitchell. Generalization as Search. Artificial Intelligence, Vol. 18, 1982, pp 203-226.

....control, namely the use of syntactic rule models, and the use of a task oriented domain topology. We briefly describe some preliminary application results of RDT within the knowledge acquisition system MOBAL, and present directions of further research. 1 Introduction As described by Mitchell [Mit82] concept learning from examples can be seen as a process of search in a space of hypotheses that is ordered by the generality relationship between hypotheses. The size of this space and the properties of its generalization relation determine the difficulty of finding the desired target concept, ....

Tom M. Mitchell. Generalization as search. Artificial Intelligence, 18(2):203 -- 226, 1982.

....incremental. In nonincremental mode, the learning algorithm infers a model based on an entire set of available training samples (e.g. ID3 in [51] In incremental mode, the learning algorithm revises a model in response to each newly observed measurement (e.g. Candidate Elimination Algorithm in [45]) An incremental algorithm may be appropriate for learning tasks in which a stream of measurements is available (see [63] In feeding 3D parts, incremental approaches may be more appropriate for this task, on the assumption that the task is a serial learning task and the cost of computing plans ....

Tom M. Mitchell. Generalization as search. Artificial Intelligence, 18:203-- 226, 1982.

....to seeking a concept description which characterizes all the positive examples of the concept while specialization consists of seeking a concept description which excludes all the counter examples. This framework is illustrated in Figure 1. 2 and it will be formalized in Chapter 2, following [Mitchell1982] in using the notion of Version Spaces . Within this generalization specialization framework, discrimination based classifiers learn concepts by generalizing from the examples of the concept while specializing using the counter examples. Although recognition based learners are also able to ....

....not, however, have access to counter examples for specialization. Assuming this framework, an important question then becomes: how can inductive systems learn how to discriminate between the concept and the counter concept (or complementary concept) class in the absence of counter examples Since [Mitchell1982] s framework suggests that, in the absence of counter examples, an inductive learning system will never stop 5 generalizing (because it is not aware of any boundaries) a recognition based system is expected to over generalize the concept class and thus be an inappropriate classification tool. ....

[Article contains additional citation context not shown here]

Tom Michael Mitchell. Generalization as search. Artificial Intelligence, 18(2):203--226, March 1982.

....method that commences with greedy search, extending its scope at each iteration until a stopping criterion is satisfied. This layered search is often found to produce theories that are more accurate than those obtained with either greedy search or moderately extensive beam search. 1 Introduction Mitchell [1982] observes that the generalization implicit in learning from examples can be viewed as a search over the space of possible theories. From this perspective, most machine learning methods carry out a series of local searches in the vicinity of the current theory, selecting at each step the most ....

Tom Mitchell. Generalization as search. Artificial Intelligence 18, 203-226.

....traces, and then refines these through exploration. The set of operator preconditions is represented as two sets, one a most specific version and the other a most general version, consist with the observed training instances. These sets are refined in a manner similar to that of version spaces [Mitchell, 1982] and relies on operator preconditions being a single, conjunctive set. The STRIPS like planning representation allows these systems to reason in large state and goal spaces (E1and E2) and (for EXPO and OBSERVER) to use a full planner from the PRODIGY architecture [Minton et al. 1989; Carbonell et ....

Tom M. Mitchell. Generalization as search. Artificial Intelligence, 18(2):203--226, 1982.

....of the threshold. The cycle is described in Figure 1 for the finite case. Single classifier approaches to uncertainty sampling have been criticized [6, 20] on the grounds that one classifier is not representative of the set of all classifiers consistent with the labeled data: the version space [24]. The degree to which this is a problem in practice has not been established. Single classifier approaches have successfully been used in generating arbitrary queries [16] and in sampling from labeled data [8, 25] Uncertainty sampling with a single classifier can also be viewed as a variation on ....

Tom M. Mitchell. Generalization as search. Artificial Intelligence, 18:203--226, 1982.

....language that are more general than constant values. The classical approaches to overcome this problem are hardly applicable when learning from real databases. First, taxonomies of descriptors have served this purpose, for instance, in conceptual clustering [14] or the versions space approach [16]. There, however, the taxonomies were given by the user of the learning system. Here, we do not know of such hierarchies of sets of attribute values but have to find them. Second, methods for constructive induction exist that introduce more abstract descriptors into the hypothesis language, but ....

Tom M. Mitchell. Generalization as search. Artificial Intelligence, 18(2):203-- 226, 1982.

....some stimulus may be non deterministic, i.e. relational. In most positions on a chess board, there are reasonable alternative moves. To most questions, there is more than one answer, even writing this paper could be continued in several ways. This corresponds to the version space algorithm (see [Mit82] because classification based on incomplete information possesses some learning aspect. 3 VALIDATION SCENARIOS Every approach to system validation by testing needs to know what questions to ask . It is the aim of our present paper to identify those questions by exploiting structural ....

Tom M. Mitchell. Generalization as search. Artificial Intelligence, 18:203--226, 1982.

....values that appear across multiple examples. Given enough examples, the correlations indicate which features are important to the concept being learned. Correlational techniques include symbolic concept learning and clustering methods (e.g. Fisher, 1987; Quinlan, 1986; Michalski and Stepp, 1983; Mitchell, 1982] connectionist methods (e.g. Rumelhart and McClelland, 1986] and other reinforcement learning techniques (e.g. Sutton, 1988; Holland, 1986] Since instructions may be explicitly situated, determining the situation for an instruction would have to be included as a precursor to ....

.... remembers a single version of each piece of knowledge, inability to recover from incorrect knowledge rules out learning from multiple examples, which would require changing the single hypothesis (an alternative would be to explicitly represent multiple hypotheses; e.g. in a version space [Mitchell, 1982]) Similarly, it precludes instruction by general case and exceptions; for instance, Never grasp red blocks, and then later, It s ok to grasp the ones with safety signs on them. 1 Recovery from incorrect knowledge is discussed further below as one component of a specific proposal for future ....

Tom M. Mitchell. Generalization as search. Artificial Intelligence, 18(2):203--226, 1982.

....in rule based systems. Since this is a relative distinction, categorizing the general CHAPTER 1. INTRODUCTION 7 level of knowledge is done by defining a more general than or more specific than relation. An implementation independent definition for these relations can be described as below [ Mitchell, 1982 ] Rule r 1 is more general than rule r 2 (or r 2 is more specific than r 1 ) if in any world r 1 can be used to show at least the same results as r 2 . We have found three different but related approaches to rule generality in which the definition of more general than or more specific than ....

....according to (2) which is also more specific than A(x; b; z) B(v; w) G according to (3) where v; w; x; z are variables. The second definition of rule generality is used in concept learning where a general description of a class of objects is derived from a set of examples and non examples [ Mitchell, 1982 ] In this work, each instance is described by an unordered pair of feature vectors, each CHAPTER 1. INTRODUCTION 8 of which specifies the size, color, and shape of an object. For example, f(Large Red Square) Small Yellow Circle)g is an instance. Generalizations of these instances are ....

Tom M. Mitchell. Generalization as search. Artificial Intelligence, 18:203--226, 1982.

....by any rule classifier, accumulate it into a dataset. If the number of correlations of the same type (i.e. same grammar 7 A rule classifier is applicable to a second order correlation if its grammar and control components match those of the correlation. 8 Unlike the version space algorithm [8], our algorithm does not maintain all the most general and most specific generalizations consistent with the current set of examples. Our algorithm can also handle disjunctive generalizations and noise. bits, and same shift and unlock actions) exceeds a threshold, invoke the summarization ....

Tom Mitchell. Generalization as search. Artificial Intelligence Journal, 18, 1982.

....If we suppose the current situation includes fgoal climb rate = 1 5, current climbrate = 0, and heading = Northg, this situation would map to the sample case if there were no more specific situation which also matched the current situation. Casting this procedure as a problem space search [10], there are two opera12 tors, remove a parameter and generalize a parameter, which can apply to the state, the current situation. The goal is to reach a state matching a generalized situation in the case database in the fewest possible steps. Search control in this problem space is ....

Tom M. Mitchell. Generalization as search. Artificial Intelligence, 18(2):203--226, 1982.

.... on a first order formulation [Russell and Grosof, 1987] of the biases in the Version Space method [Mitchell, 1978] Implementation of inference in the non monotonic logic then enables the principled, automatic modification of the description space employed in a concept learning program, which Mitchell [1982] remarked would represent a significant advance in this field , and which Bundy et al. 1985] named as the most urgent problem facing automatic learning . We also showed how to formulate with prioritized defaults two kinds of preference biases previously regarded as syntactic or ....

Tom M. Mitchell. Generalization as search. Artificial Intelligence, 18(2):203--226, 1982.

....we expect that the MLP network should be better suited than the autoassociator to learning domains that require a strong specialization bias caused by the conceptual class, provided that the counter conceptual class contains meaningful specialization information. 2 These ideas are formalized in [Mitchell1982] using the notion of Version Spaces . This is because, the MLP network has the possibility to rely on the counter examples during the inductive process, whereas the autoassociator does not. On the contrary, if domain requirements for a strong specialization come from the counter conceptual ....

Tom Michael Mitchell. Generalization as search. Artificial Intelligence, 18(2):203-- 226, March 1982.

....alternates between generating an exploration plan, learning from the results of the exploration, and attempting straight problem solving. The learning phase involves a fairly straightforward generation and specialization process, and can be thought of in terms of Mitchell s version spaces [10] and related work. The problem solving phase is simple STRIPS style planning. The novelty lies in the creation of the exploration plans, which are intended to provide suprising experiences to serve as the basis for new rules. This process is reminiscent of the creation of practice problems in ....

Tom M. Mitchaell. Generalization as search. In Shavlik and Dietterich [17], pages 96--107.

No context found.

Tom Mitchell. Generalization as Search. Artificial Intelligence, 18(2):203--226, 1982.

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