R.M. Keller T.M. Mitchell and S.T. Kedar-Cabelli. Explanation-based learning: A unifying view. Machine Learning, 1(1):47--80, 1986.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....in later sessions to provide explanations associated with transitions between problem solving stages. The creation of explanations for events is a complex problem, especially when the knowledge is incomplete, not explicit (as part of the DM s mental model of the decision situation) or inaccurate [10]. Though the problem of generating explanations has not been solved in general, several CBR systems [6] use inbuilt domain knowledge and causal models to generate limited explanations. When the generation of explanations is hard, CBR systems [6] frequently turn to the DM for help in creating ....

Mitchell, R.M., Kellar, R.M., and Kedar-Babelli, S.T. Explanation-based learning: A unifying view. Machine Learning 1, 1. (1986), pp. 47-80.

....thousands of extra nodes without finding better quality plans than those produced by the search control system. For all that work, no improvement in quality 5 Related Work The basic idea of learning search control rules to speed up problem solving can be traced back to the early work on EBL [11, 10]. Minton s [10] PRODIGY EBL learned control rules by explaining why a search node leads to success or failure. Kambhampati et al. 6] propose a technique based on EBL to learn control rules for partial order planners and apply it to SNLP and UCPOP to learn rejectionrules. Ihrig et al. 4] extended ....

T. Mitchell, R. Keller, and S. Keddar-Cabelli. Explanation based learning: A unifying view. Machine Learning, 1:47--80, 1986.

....searches in the abstract plan space containing only static predicates and dynamic predicates owned by the sort s; b) grounds the task configuration with typical instances of sorts, if necessary. Step 6. 2 generalises the abstract plan along similar lines to a standard explanation based technique [55]: the weakest precondition of the abstract plan is assembled, and all instances of sorts that are unified with variables in the preconditions of operators during execution are carefully generalised to variables in the final plan. Returning to our example above, we have the generalised task ....

T. M. Mitchell and R. Keller and Kedar-Cabelli. Explanation--Based Learning: a unifying view. Machine Learning, 1, 1986.

....to. Second, and perhaps more ambitious, it can also learn to avoid similar failures in the future. Both these capabilities can be provided by the general analysis of explanations of the failures encountered by the planner. Explanation based learning techniques (EBL) studied in machine learning [35,10,32], offer significant promise in this direction. The general idea behind explanation based learning (see Figure 1) is as follows: given a problem the planner searches through the space of possible solutions and returns a solution. The learner then explains the failures and successes in the search ....

.... are variablized (the specific instances are replaced by fresh copies of the corresponding operation schemas) the proved target concept is variablized, and regressed through the generalized proof tree to compute the weakest conditions under which the variablized target concept can be proved again [31,35,45]. In the context of SNLP EBL, the proof tree is the part of the search tree that terminates in failing nodes, and the operations are the planner decisions, and the generalization process will involve variablizing the planner decisions in the failing search tree, starting with variablized failure ....

[Article contains additional citation context not shown here]

T. M. Mitchell, R. M. Keller, and S. T. Kedar-Cabelli. Explanation-based learning: A unifying view. Machine Learning, 1(1):47 -- 80, 1986.

.... beliefs b added also to the set D x y as d: b] A test of consistency will be run on TLI x y and, in case of 9 inconsistency, x will remove the inconsistency by imposing a preference relation x y relation over Part(D x y ) The approach resembles the first phase of an EBL procedure [10] where the example is represented by the elements of Base(D xy ) We are currently investigating: an efficient method for establishing preferences on Part(D x y ) the possibility to apply the second phase of an EBL procedure to generalize the plausible model of y, and the efficient revision of ....

T.M. Mitchell, R.M. Keller, and S.T. Kedar-Cabelli, Explanation-based learning: A unifying view, Machine Learning 1 (1986) 47-80.

....as a set of causal patterns. A total of approximately 25 causal patterns have been implemented. The causal patterns make use of temporal and spatial constraints to suggest causal relationships. One such causal pattern is given below: 3. A theory of causation is also termed a domain theory (Mitchell, Kedar Cabelli Keller 1986). 8 When similar actions on an object are followed by a state change, and a feature of the object correlates with the state change, then that feature is need for the action to result in the state change. In TDL, a causal pattern that matches a training example proposes a hypothesis which is ....

Mitchell, T., Kedar-Cabelli, S. & Keller, R. (1986). Explanation-based learning: A unifying view. Machine Learning, Vol. 1(1).

....each clause, FOIL removes from further consideration the positive examples covered by that clause. The learning process ends when all positive examples have been covered by some clause. FOCL (Pazzani Kibler, 1991) extends FOIL by incorporating a compatible explanation based learning component (Mitchell, et al. 1986). This allows FOCL to take advantage of existing knowledge provided by experts. When constructing a clause body, there are two ways that FOCL can add predicates. First, it can add predicates via the same inductive method used by FOIL. Second, it can add predicates by deriving them from the rule ....

Mitchell, T., Keller, R., & Kedar-Cabelli, S. (1986). Explanation-based learning: A unifying view. Machine Learning, 1.

....it is necessary to index the case in more than one manner and to include a description of the encoding context in the index. The creation of indices by explanation based techniques can accomplish this by explaining more than one aspect of a case. At storage time, several different target concepts [7] can focus the learning process on explaining different parts of an episode. For example, to predict the outcome of future incidents it is necessary to explain why previous incidents fail or succeed. By also explaining why the actor tried economic sanctions rather than some other action (e.g. ....

Mitchell, T., Kedar-Cabelli, S. & Keller, R. (1986). Explanation-based learning: A unifying view. Machine Learning, Vol. 1(1).

.... in the logical consistency group learn faster and make fewer errors than subjects in the feature consistency group, they learner slower and make more errors than would be predicted by existing computational models of the influence of prior knowledge such as Explanation based learning (EBL) (Mitchell et al. 1986). EBL is a machine learning method that derives concepts from background knowledge. At first, it might seem that EBL would serve as an ideal model of the use of prior knowledge in learning. Its inputs correspond exactly to those items learned in Phases 1 3 of the first experiment, and its output ....

Mitchell, T., Keller, R., & Kedar-Cabelli, S. (1986). Explanation-based learning: A unifying view. Machine Learning, Vol. 1(1).

....Several systems have been constructed around the idea of learning control rules. These systems are based around some search (or problem solving) method and various techniques are used to acquire control knowledge that can direct the search. The techniques include Explanation Based Learning methods [11, 29] used in Prodigy [27] and in Soar [24, 34] static analysis in Prodigy [14] analogy in Prodigy [45, 46, 47] and Inductive Logic Programming in Scope [13] Our approach clearly differs in the method of acquiring rules, but perhaps more importantly, the strategy that our learning algorithm finds, ....

T. Mitchell, R. Keller, and S. Kedar-Cabelli. Explanation based learning: A unifying view. Machine Learning, 1:47--80, 1986.

....search control 14900 174 156 126 54 Table 4: Performance data for the logistics domain. produces in a time comparable to that required by Sys SEARCH CONTROL. 5 Related Work The basic idea of learning search control rules to speed up problem solving can be traced back to the early work on EBL [13, 12, 3]. Minton s [12] PRODIGY EBL learned control rules by explaining why a search node lead to success or failure. Kambhampati et al. 7] propose a technique based on EBL to learn control rules for partial order planners and apply it to SNLP and UCPOP to learn rejection rules. Rejection type rules are ....

T. Mitchell, R. Keller, and S. Keddar-Cabelli. Explanation based learning: A unifying view. Machine Learning, 1:47--80, 1986.

....recognition. Pure induction methods like C4.5 [15] while flexible and produce understandable domain theories, don t capitalize on prior knowledge. As a result, they require the learning process to start from scratch with large training sets. Explanation based learning (EBL) methods developed by [3,10] capitalize on prior knowledge and are understandable, but lack the flexibility for learning new correlations autonomously. The TRIPLE system, developed by Ming and Bhanu [9] is the most notable EBL system that has been used successfully for automatic target recognition. The classifier in TRIPLE ....

Mitchell, T.M., Keller, R.M., and Kedar-Cabelli, S.T., "Explanation-Based Learning: A Unifying View", Machine Learning, vol. 1, no. 1, 1986, pp. 47-80.

....that are causally related. Waldmann and Holyoak (1990) argue that the causal induction process differs from the learning process used to acquire arbitrary concepts. While we have been inspired by work in machine learning on analytic learning algorithms such as explanationbased learning (EBL) (Mitchell et al. 1986, Mooney DeJong 1986) it is obvious that our human subjects cannot be modeled by existing EBL algorithms. In particular, EBL algorithms would learn more quickly than the logical consistency subjects of Experiment 1. Since the fourth concept can be deductively derived from the preceding three, ....

Mitchell, T., Keller, R., & Kedar-Cabelli, S. (1986). Explanation-based learning: A unifying view. Machine Learning, Vol.

....learning within each component and to control the interaction between components. FOCL was designed to learn Horn clause descriptions by combining analytical and empirical learning. As a consequence of this combination of learning methods, FOCL can utilize incomplete and incorrect domain theories (Mitchell, Keller, Kedar Cabelli, 1986; Rajamoney, DeJong, 1987) FOCL uses the rule base of an expert system as the domain theory required by explanation based learning and processes a collection of examples. FOCL extracts the portions of the domain theory that correctly classify examples, patches the portions of the domain theory ....

Mitchell, T., Keller, R., & Kedar-Cabelli, S. (1986). Explanation-based learning: A unifying view. Machine Learning, 1.

....the goal through the explanation structure using a modified version of the goal regression algorithm described by Waldinger [27] and Nilsson [23] The conjunction of the resulting expressions determine the conditions under which the generalized plan can be used to achieve the goal. See [9] [20], and [25] for further details. These two steps can be summarized as follows: The first step creates an explanation structure that determines the relevant steps in a sequence of operations that are necessary to achieve the goal. The second step analyses this explanation to determine the ....

....sufficient conditions and operations for achieving a set of postconditions, and uses the input example to focus the search for an explanation. As a result, if the input sequence is not an optimal way to achieve a postcondition, the algorithm would not be able to detect this. As pointed in [20], the use of an input example to focus the search for an explanation is desirable in order to formulate relevant explanations. However, we believe that a useful extension of the algorithm would be to determine an optimal explanation (using the fewest number of steps and or fewest unsupported ....

Tom M. Mitchell, Richard M. Keller, and Smadar T. Kedar-Cabelli. Explanation-based learning: A unifying view. Machine Learning, 1:47--80, 1986.

....feature values may result in completely different values being similar to some extent. This is due to propagation of similarities through the graphs. 6 Numerical values are compared by employing a sigmoid function. 6. 1 Modified Explanation Based Learning Since pure explanation based learning ([10, 11, 12]) seams not applicable 7 , a slightly modified strategy is applied: Similar to [13] the user is asked for matching knowledge when a diagnosis is finally selected for a case and discrepancies concerning that diagnosis are detected (i.e. a user defined constraint is violated or there are other ....

T. M. Mitchell, R. M. Keller, and S. T. Kedar-Cabelli. Explanation-based learning: A unifying view. In J. W. Shavlik and T. G. Dietterich, editors, Readings in Machine Learning. Morgan Kaufmann, San Mateo, 1990.

....of extensions allows FOCL to use intensionally defined predicates (i.e. predicates defined by a rule instead of a collection of examples) in a manner similar to the extensionally defined predicates in FOIL. A collection of intentionally defined predicates is identical to the domain theory of EBL (Mitchell, Keller, Kedar Cabelli, 1986). A final extension allows FOCL to accept as input a partial, possibly incorrect rule that is an initial approximation of the predicate to be learned. If this rule is defined in terms of extensionally defined predicates, it is analogous to a partial concept definition constructed by an incremental ....

Mitchell, T., Keller, R., & Kedar-Cabelli, S. (1986). Explanation-based learning: A unifying view. Machine Learning, 1, 47--80.

....the range of values for a particular feature is too broad and must be narrowed. Generating useful explanations is notoriously difficult. The two most common ways to generate explanations are operator assisted explanation based learning (OAEBL) 24] and explanation based generalization (EBG) [18]. In OAEBL, a knowledge engineer generates explanations. This requires an expert in the problem domain to be present for the entire training period of the EBL system. In addition, it requires a man machine interface that is easily understood by the human operator and the EBL system. As noted ....

Mitchell, T.M., Keller, R.M., and Kedar-Cabelli, S.T., "Explanation-Based Learning: A Unifying View", Machine Learning, vol. 1, no. 1, 1986, pp. 47-80.

....abstraction and decomposition knowledge that speeds up the performance of hillclimbing search algorithms. It thus automates a task that is analogous to tasks performed by some well known methods of speedup learning, such as problem reformulation [Amarel, 1968] and explanation based learning, [Mitchell et al. 1986, Laird et al. 1987] 3 Related Work Several investigators have previously applied AI techniques to parameter design optimization: Methods of intelligently adapting the parameters of hillclimbing search during a single Figure 1: The Stars and Stripes 87 problem solving session are reported ....

T. M. Mitchell, R. M. Keller, and S. T. Kedar-Cabelli. Explanationbased learning: A unifying view. Machine Learning, 1(1):47 -- 80, 1986.

....training data is very scarce and noisy, yet there exist much refined, though still very approximate, analytic models that have been applied for the past decades and embody many years of experience in this particular domain. Much in the spirit of Explanation Based Learning (see, for example, Mitchell et al. 1986; Minton et al. 1986) where domain knowledge is applied to get valid generalizations from only a few training examples, we consider an analytic model as an imperfect domain theory from which the training data is explained (see also Scott et al. 1991; Bergadano et al. 1990; Tecuci et al. ....

Mitchell, T. M., R. M. Keller and S. T. Kedar-Cabelli (1986). Explanation-based Learning: A unifying view. Machine Learning, Vol. 1, pp. 47-80.

....by looking for similarities and differences between positive and negative examples of a concept. Current connectionist learning techniques (e.g. Rumelhart, Hinton, Williams, 1986) are essentially empirical learning techniques. Explanation based learning (EBL) techniques (DeJong Mooney, 1986; Mitchell, Kedar Cabelli, Keller, 1986) operate by forming a generalization from a single training example by proving that the training example is an instance of the concept. The proof is constructed by an inference process that makes use of a domain theory, a set of facts and logical implications. In explanation based learning, a ....

Mitchell, T., Kedar-Cabelli, S., & Keller, R. (1986). Explanation-based learning: A unifying view. Machine Learning, 1, 47-80.

....in the plan. Generalized plans are of particular utility in plan reuse as storing and reusing generalized plans as against specific plans allows for storage compactions, and retrieval efficiency. An important approach for plan generalization is the so called explanation based generalization (EBG) [14]. Much of the previous work on EBG of plans concentrated on inflexible plan representations, and was limited to straight forward precondition and order generalization [14, 18, 10, 2, 17, 19, 3] This paper aims to unify and extend the previous work on plan generalizations. We start with the ....

....and retrieval efficiency. An important approach for plan generalization is the so called explanation based generalization (EBG) 14] Much of the previous work on EBG of plans concentrated on inflexible plan representations, and was limited to straight forward precondition and order generalization [14, 18, 10, 2, 17, 19, 3] This paper aims to unify and extend the previous work on plan generalizations. We start with the flexible partial order plan representation, and formalize explanation based plan generalization in terms of proofs of correctness with respect to modal truth criteria (which have been originally ....

T. M. Mitchell, R. M. Keller, and S. T. Kedar-Cabelli. Explanation-based learning: A unifying view. Machine Learning, 1(1):47 -- 80, 1986.

No context found.

R.M. Keller T.M. Mitchell and S.T. Kedar-Cabelli. Explanation-based learning: A unifying view. Machine Learning, 1(1):47--80, 1986.

No context found.

T. M. Mitchell, R. M. Keller, and S. T. Kedar-Cabelli. Explanation-based learning: A unifying view. Machine Learning, 1(1):47 -- 80, 1986.

No context found.

T. M. Mitchell, R. M. Keller, and S. T. Kedar-Cabelli. Explanation-based learning: A unifying view. Machine Learning, 1(1):47 -- 80, 1986.

First 50 documents

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