Steven Minton. Learning Effective Search Control Knowledge: An Explanation-Based Approach. PhD thesis, Computer Science Department, Carnegie Mellon University, 1988. 194  Home/Search   Document Not in Database   Summary   Related Articles   Check

....in which knowledge gaps are identified, the reasons why particular hypotheses are generated, the strategies chosen for verifying candidate hypotheses, and the basis for choosing particular reasoning methods for each of these. Trace Meta XPs are similar to reasoning traces [Carbonell, 1986; Minton, 1988; Veloso Carbonell, 1993] or justification structures [Collins, Birnbaum, Krulwich, Freed , 1993; deKleer, Doyle, Steele, Sussman , 1977; Doyle, 1979] with the difference that Trace Meta XPs represent, in addition to the subgoal structure of the problem and justifications for operator ....

S. Minton. Learning effective search control knowledge: An explanation-based approach. Ph.D. thesis, Carnegie-Mellon University, Computer Science Department, Pittsburgh, PA, 1988. Technical Report CMU-CS-88-133.

....actually improve performance according to any realistic metric is, at best, inconclusive. For example, in [ Tambe and Newell, 1988 ] it is shown that chunking in SOAR can either degrade or improve performance, depending on the domain; and independent analyses in [ Shavlik, 1987 ] and [ Minton, 1988a ] reach contradictory conclusions about the asymptotic behavior of explanation based learning in improving performance of planners in the blocks world domain. In fact, it is hardly surprising that experimental evidence has failed to completely settle this issue, given the time requirements ....

....investigated, the questions when is SLL learning useful and even is SLL useful will not have been answered. In closing, we would like to make an attempt to dispell one possible source of confusion. Readers familiar with chunking and EBL techniques as used in problem solvers like Prodigy [ Minton, 1988a ] or SOAR [ Laird et al. 1986 ] may be aware that in these systems, chunking serves mostly to improve the control decisions made by the problem solver in particular to reorder goals and resolve goal conflicts. It may appear at first that the techniques in this paper are quite different; in ....

[Article contains additional citation context not shown here]

Steven Minton. Learning effective search control knowledge: An explanation-based approach. Technical report, Carnegie-Mellon University Department of Computer Science, 1988.

....formalize the reasons that this is so. What s wrong with Strawman 1 Simply put, Strawman 1 s behavior is unsatisfying because its outputs are too general. In most EBL contexts, there is a penalty for overgeneralizing an example. If the learned rules are used to speed up problem solving, as in [Minton, 1988], then they will be overgeneral and (in most cases) needlessly inefficient; if the learned One possible choice for R is a recursive theory, as in [ Shavlik, 1990 ] rules are used to construct a new theory that better models the data, as in [Cohen, 1990b] then they will be inaccurate. How ....

Steven Minton. Learning effective search control knowledge: An explanationbased approach. Technical report, Carnegie-Mellon University Department of Computer Science, 1988.

....i.e. to automate the acquisition of knowledge that guides the problem solving search process. One approach to learning control knowledge consists of generating explanations for the local decisions made during the search process (DeJong and Mooney, 1986, Laird et al. 1986, Mitchell et al. 1986, Minton, 1988, Prez and Etzioni, 1992, Katukam and Kambhampati, 1994) These explanations become control rules that are used in future situations to prune the search space. These deductive approaches invest a substantial explanation effort to produce proven correct and complete control rules from a single (or ....

....of the search tree; Credit assignment; and Generation of control rules. The Bounded Explanation module behaves lazily in two aspects: It does not require to learn initially correct or complete control knowledge. In contrast with other eager approaches for learning control knowledge for planning (Minton, 1988, Etzioni, 1993) HAMLET does not require the learned knowledge to be correct initially. The incremental refinement will be responsible for the correctness of the control knowledge at the end of the learning process. Therefore, there is no need for additional domain axioms. It does not require to ....

[Article contains additional citation context not shown here]

Steven Minton. Learning Effective Search Control Knowledge: An Explanation- Based Approach. Kluwer Academic Publishers, Boston, MA, 1988.

....Prodigy I Introduction Prodigy is an integrated planning and learning system. The system includes not only a planning algorithm but also procedures for learning and case based reasoning, which greatly increase the efficiency of the planner. For example, Prodigy is able to learn control rules [Minton, 1988], conduct experiments to acquire new knowledge [Gil, 1992] generate abstraction hierarchies [Knoblock, 1993] and use andlogical reasoning to recognize and exploit similarities between planning problems [Veloso, 1992] Prodigy s core, the planning algorithm itself, has been improved over the ....

Steven Minton. Learning effective search control knowledge: an explanation- based approach. Kluwer Academic Publishers, Boston, MA, 1988. 10

....Anderson and Farley, 1990] In PRODIGY, there is a clear division between the declarative domain knowledge (operators and inference rules) and the more procedural control knowledge. This simplifies both the initial specification of a domain and the incremental learning of the control knowledge [Minton, 1988, Veloso and Carbonell, 1990] Previous work in the linear planner of PRODIGY used explanation based learning techniques [Minton, 1988] to extract from a problem solving trace the explanation chain responsible for a success or failure and compile search control rules. We are now extending this ....

....rules) and the more procedural control knowledge. This simplifies both the initial specification of a domain and the incremental learning of the control knowledge [Minton, 1988, Veloso and Carbonell, 1990] Previous work in the linear planner of PRODIGY used explanation based learning techniques [Minton, 1988] to extract from a problem solving trace the explanation chain responsible for a success or failure and compile search control rules. We are now extending this work to NOLIMIT, as well as developing a derivational analogy approach to acquire control knowledge [Carbonell, 1986, Veloso and ....

Steven Minton. Learning Effective Search Control Knowledge: An Explanation-BasedApproach. Kluwer Academic Publishers, Boston, MA, 1988.

....which supports our research goals underlying hamlet s learning algorithm. The overall running times also decreased using the rules, but not significantly. We did not find empirically with our learned rules that the time spent solving the problem degraded so much to consider it a utility problem [11]. However, we are currently developing efficient methods for organizing and matching the learned control rules. We consider this organization essential and part of the overall learning process [4] 5 Related work Most speedup learning systems have been applied to problem solvers with the ....

.... [4] 5 Related work Most speedup learning systems have been applied to problem solvers with the linearity assumption, such as the ones applied to Prolog or logic programming problem solvers [15, 21] special purpose problem solvers [12, 9, 18] or other general purpose linear problem solvers [5, 10, 11, 14]. These problem solvers are known to be incomplete and and incapable of finding optimal solutions. If we remove the linearity assumption, we are dealing with nonlinear problem solvers. This kind of problem solvers are needed to address real world complex problems. In general, there have not been ....

[Article contains additional citation context not shown here]

Steven Minton. Learning Effective Search Control Knowledge: An Explanation-Based Approach. Kluwer Academic Publishers, Boston, MA, 1988.

....for solving efficiently any problem of the domain. Two approaches adopted have been to learn an explanation and prove that it is correct, or not prove that it is correct and refine it upon experiencing on other problems. The first approach, as in [ DeJong and Mooney, 1986, Mitchell et al. 1986, Minton, 1988 ] usually involves a substantial effort to prove the correctness of the learned knowledge. In addition it requires a complete domain theory to obtain the explanations , although there have some work on learning with incomplete, or intractable theories, such as [ Tadepalli, 1989 ] Moreover, ....

....can select, reject, prefer, or decide on the choice of alternatives [ Veloso, 1989 ] This knowledge guides the search process and helps to reduce the exponential explosion in the size of the search space. Previous work in the linear planner of prodigy uses explanation based learning techniques [ Minton, 1988 ] to extract from a problem solving trace the explanation chain responsible for a success or failure and compile search control rules therefrom. Similar efforts within the linear planner of prodigy were done to learn control rules from partially evaluating the domain theory [ Etzioni, 1990, ....

Steven Minton. Learning Effective Search Control Knowledge: An ExplanationBased Approach. PhD thesis, Computer Science Department, Carnegie Mellon University, 1988. Available as technical report CMU-CS-88-133.

....should be modified. The second computes the appropriate modification for each identified heuristic. The final part installs the modification into the problem solver s set of explicit search control rules. A restricted version of this algorithm has been implemented as an extension to the PRODIGY[13] problem solver. This extension is called Bacall. The main restriction is that Bacall can only modify the linearity[19] and the strong linearity [6] rejection heuristics. Linearity directs the problem solver to work on goals in a strict depth first fashion, i.e. pick a toplevel goal and an ....

....between the quality of the rejection rules learned and the cost of learning them. HAMLET is an example of a system that attempts such a compromise. While these systems learned more than just rejection search control rules, I will only discuss that one aspect. 2.2. 1 PRODIGY EBL PRODIGY EBL[13] learned rejection search control rules from failed search subtrees. The learner used a theory of the problem solver to identify the leaf failure reasons which were then regressed up the subtree. While the theory was relatively complete, it did not capture all of the relevant aspects of every ....

[Article contains additional citation context not shown here]

S. Minton. Learning Effective Search Control Knowledge: An Explanation-Based Approach. Kluwer Academic Publishers, Boston, MA, 1988.

....guiding heuristics. One approach to learning control knowledge from a problem solving trace consists of generating explanations for the individual decisions made during the search process. These explanations become control strategies that are used in future situations to prune the search space [16]. There is also work done on doing the generation of control rules without problem solving episodes, by statically looking at the domain description [8] However, these strong deductive approaches invest a substantial explanation effort to produce correct control strategies from a single problem ....

....learned control knowledge to guide the search and convert it into an intelligent commitment search strategy [22] Control knowledge guides the search process and helps to prune the exponential search space. Previous work in the linear planner of prodigy uses explanation based learning techniques [16] to extract from a problem solving trace the explanation chain responsible for a success or failure and compile search control rules therefrom. Similar efforts within the linear planner of prodigy were done to learn control rules from partially evaluating the domain theory [8, 19] The paper ....

[Article contains additional citation context not shown here]

Steven Minton. Learning Effective Search Control Knowledge: An ExplanationBased Approach. PhD thesis, Computer Science Department, Carnegie Mellon University, 1988. Available as technical report CMU-CS-88-133.

....approaches. We hope to use ideas from Soar to address structure learning in Icarus. Prodigy [23] is an architecture for integrating planning and learning. Like Soar, it has been applied to execution systems, but the main emphasis has been on problem solving. An early version of Prodigy [12] applied explanation based learning to develop new rules that govern which states to expand or operators to select. More recent work adds learning methods with mutually interpretable knowledge structures, such as analogical reasoning and learning by experimentation. These methods support the goals ....

Minton, S. (1988). Learning effective search control knowledge: An explanation-based approach. Boston, MA: Kluwer Academic Publishers,.

....operational. Explanationbased learning (EBL) approaches have been successfully used to make problem solving knowledge operational for a number of different types of problem solving including concept formation (Mitchell, Kellar, and Kedar Cabelli 1986) planning (Fikes, Hart, and Nilsson 1972; Minton 1988) and scheduling (Carbonell, Knoblock, and Minton 1991) EBL uses a domain theory to generate an explanation of why some training instance is an example of a goal concept according to some operationality criterion (DeJong and Mooney 1986) In external domains, the goal concept for EBL is the ....

....the result of compilation includes not only Rule 6 but this rule as well: 7. IF Goal(Put On Table(x) Clear(x) Left Of(Gripper, x) Higher(Gripper, x) THEN Execute(Step(right, Gripper) A commitment to a particular level of operationality may be viewed as a way to address the utility problem (Minton 1988). If the utility problem may be ignored (due to the nature of the environment or the characteristics or the problem solver) then the best level of operationality may be determined through performance, rather than making an arbitrary assumption a priori. In this, the best case, the task ....

[Article contains additional citation context not shown here]

Minton, S. (1988). Learning Effective Search Control Knowledge: An ExplanationBased Approach. Kluwer Academic Publishers.

....to produce finer grained abstraction hierarchies. The article describes these extensions in detail. 4 1. 4 Experimental Results alpine has been successfully tested on a number of planning domains including the Tower of Hanoi, the original strips domain [20] a more complex robot planning domain [44], and a machine shop process planning and scheduling domain [44] In all these domains, the system generates problem specific abstraction hierarchies that provide significant reductions in search. The algorithm for generating the abstractions is quite efficient and can generate an abstraction ....

....describes these extensions in detail. 4 1. 4 Experimental Results alpine has been successfully tested on a number of planning domains including the Tower of Hanoi, the original strips domain [20] a more complex robot planning domain [44] and a machine shop process planning and scheduling domain [44]. In all these domains, the system generates problem specific abstraction hierarchies that provide significant reductions in search. The algorithm for generating the abstractions is quite efficient and can generate an abstraction hierarchy for a problem in any of these domains in 0.3 to 4.5 CPU ....

[Article contains additional citation context not shown here]

Steven Minton. Learning Effective Search Control Knowledge: An Explanation-Based Approach. PhD thesis, Computer Science Department, Carnegie Mellon University, 1988.

....Anderson and Farley, 1990] In PRODIGY, there is a clear division between the declarative domain knowledge (operators and inference rules) and the more procedural control knowledge. This simplifies both the initial specification of a domain and the incremental learning of the control knowledge [Minton, 1988, Veloso and Carbonell, 1990] Previous work in the linear planner of PRODIGY used explanation based learning techniques [Minton, 1988] to extract from a problem solving trace the explanation chain responsible for a success or failure and compile search control rules. We are now extending this ....

....rules) and the more procedural control knowledge. This simplifies both the initial specification of a domain and the incremental learning of the control knowledge [Minton, 1988, Veloso and Carbonell, 1990] Previous work in the linear planner of PRODIGY used explanation based learning techniques [Minton, 1988] to extract from a problem solving trace the explanation chain responsible for a success or failure and compile search control rules. We are now extending this work to NOLIMIT, as well as developing a derivational analogy approach to acquire control knowledge [Carbonell, 1986, Veloso and ....

Steven Minton. Learning Effective Search Control Knowledge: An Explanation-BasedApproach. Kluwer Academic Publishers, Boston, MA, 1988.

....planner may thus be improved by the use of these learned rules. Although there has been a considerable amount of research towards applying EBL techniques to planning, almost all of it concentrated on the state based models of planning, as against the partial order (plan space) models of planning [31,2]. One of the reasons for the concentration of explanation based learning (EBL) work on state space planners has been the concern that a sophisticated planner may make the learning component s job more difficult (c.f. 32] However, given the current status of plan space planning as the dominant ....

.... 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]

S. Minton. Learning Effective Search Control Knowledge: An Explanation- Based Approach. PhD thesis, Carnegie-Mellon University, Pittsburgh, PA, 1988.

....from which to learn are unambiguously provided in the input representation (Wisniewski Medin, 1991) 3. Speed up learning Let me close by addressing an issue that is not raised by Cohen, but one which I am often asked: What are the implications of the OCCAM architecture on the utility (Minton, 1988) problem After all, OCCAM acquires redundant knowledge, and it is possible that it would take more time to make a prediction by traversing memory to find a very specific complex schema than it would take for an inference process to chain together several general simple schemata. The easy answer ....

Minton, S. (1988). Learning effective search control knowledge: An explanation-based effect.

....interest as a bench mark for learning to act in dynamic,uncertain worlds, and we briefly discuss it in the next section. Machine learning techniques have been studied extensively to improve the efficiency of classical AI planning systems that do not include any representation of uncertainty (Minton 1988; Veloso 1994; Borrajo Veloso 1996; Estlin Mooney 1996; Katukam Kambhampati 1994) and these techniques are still useful for reducing search in domains with uncertainty. Some of the machine learning methods used in classical planners find new uses for the conditional plans now considered. ....

....point of providing effective generalization information. 2 Several learning algorithms applied to problem solving generate explanations for the local decisions made during the search process (e.g. Laird, Rosenbloom, Newell 1986; Mitchell, Keller, Kedar Cabelli 1986; DeJong Mooney 1986; Minton 1988; Perez Etzioni 1992; Katukam Kambhampati 1994) These explanation based techniques follow a deductive approach and invest a substantial explanation effort to produce proven correct and complete control rules from a single (or few) problem solving examples and a correct and complete ....

Minton, S. 1988. Learning Effective Search Control Knowledge: An Explanation-Based Approach. Boston, MA: Kluwer.

....decisions, which means that it must perform depth first search. We cannot use breadth first or best first search; however, breadth first search in casual commitment bidirectional planners is impractically slow anyways, due to a large branching factor. Acknowledgements PRODIGY2 was designed by Steven Minton, Jaime Carbonell, Craig Knoblock, and Dan Kuokka. The next version, NOLIMIT, was created by Manuela Veloso and Daniel Borrajo. Based on these two algorithms, Jim Blythe, Manuela Veloso, Xuemei Wang, Dan Kahn, and Alicia Perez implemented PRODIGY4. Finally, Manuela Veloso and Peter Stone designed ....

Steven Minton. Learning Effective Search Control Knowledge: An Explanation-BasedApproach. PhD thesis, School of Computer Science, Carnegie Mellon University, 1988.

....on planning efficiency, that is, on acquiring problem solving strategies that control search in order to make problem solving more efficient. This area of research has been termed speed up learning . Several techniques have been used in this framework, including learning search control knowledge [20, 19, 29, 7, 24, 17, 1, 13] , macro operators [8, 15, 3, 28] chunking [16] abstraction hierarchies [14, 5] and problem solving cases [31] This thesis looks instead at the application of machine learning to acquire strategies that guide a planner at plan generation time towards improving plan quality. This aspect of ....

....or critics take care of these goal interactions by establishing ordering constraints among the conflicting goals. In the case of PRODIGY goal preference control knowledge is automatically acquired to deal effectively (in the sense of problem solving effort) with this kind of goal interactions [19, 7, 24, 31, 1, 6] . On the other hand, quality goal interactions are not directly related to successes and failures. As a particular problem may have many different solutions, quality goal interactions may arise as the result of the particular problem solving search path explored. Figure 2 illustrates another ....

[Article contains additional citation context not shown here]

Steven Minton. Learning Effective Search Control Knowledge: An Explanation-based Approach. Kluwer Academic Publishers, Boston, MA, 1988.

....in which knowledge gaps are identified, the reasons why particular hypotheses are generated, the strategies chosen for verifying candidate hypotheses, and the basis for choosing particular reasoning methods for each of these. Trace Meta XPs are similar to reasoning traces [Carbonell, 1986; Minton, 1988; Veloso and Carbonell, 1991] or justification structures [Birnbaum, Collins, Freed, and Krulwich, 1990; deKleer, Doyle, Steele, and Sussman, 1977; Doyle, 1979] with the difference that Trace Meta XPs represent, in addition to the subgoal structure of the problem and justifications for operator ....

....algorithms for learning and introspection, as well as representational methods using which a system can represent and reason about its meta models. From the artificial intelligence point of view, our approach is similar to other approaches based on reasoning traces (e.g. Carbonell, 1986; Minton, 1988] or justification structures (e.g. Birnbaum, Collins, Freed, and Krulwich, 1990; deKleer, Doyle, Steele, and Sussman, 1977; Doyle, 1979] and to other approaches that use characterizations of reasoning failures for blame assignment and or multistrategy learning (e.g. Mooney and Ourston, ....

S. Minton. Learning Effective Search Control Knowledge: An Explanation-based Approach. Ph.D. thesis, Technical Report CMU-CS-88-133, Carnegie-Mellon University, Computer Science Department, Pittsburgh, PA, 1988.

....agent. They have to classify sensor data and to determine an adequate reaction of the system in a given context. 4. Related Work An overview of work related to the application of ML to robotics has already been given in the introduction. General well known ML system architectures are Prodigy [18], Theo [19] and Soar [16] In the Prodigy system a lot of different learning strategies are investigated. But an application to a robot system has not been published so far. Theo was applied to a mobile robot (Theo agent) that collected garbage cans. Theo s learning methods (esp. EBL) were used to ....

S. Minton. Learning effective search control knowledge: An explanation-based approach. TR CMU-CS-88-133, Carnegie Mellon University, 1988.

....to be an anytime system, as described in section 3. Since we do not make any fundamental changes to the planner, we are able to take advantage of the body of work that has been done with classical planners, such as the use of abstraction [18] machine learning to improve planning performance [21, 11, 15] and derivational analogy [27, 16] To give a feel for the type of behavior we have been able to get from our architecture, in section 4 we provide two traces of the system controlling a simulated household robot built in the Oz system [1] In section 5 we present the results of some experiments ....

Steven Minton. Learning Effective Search Control Knowledge: An Explanation-Based Approach. Kluwer, Boston, MA, 1988.

....applicable to knowledge level learning of control knowledge, because it uses a separate and explicit representation of the control knowledge. A number of systems for learning meta level control constructs have been built. Among these are MetaLEX [Keller, 1988] STATIC [Etzioni, 1992] Prodigy EBL [Minton, 1988, Minton, 1990] and LEX [Mitchell et al. 1983] These systems are built on top of meta level problem solvers. The object level of the problem solver consists of domain knowledge and operators, that change the state of the problem solver. The control knowledge is explicitly represented at the ....

S. Minton. Learning Effective Search Control Knowledge: An Explanation-- Based Approach. Kluwer, 1988.

....during problem solving, the planner may have to trade off solution quality in order to find a solution at all. The planning cost depends on both the time and the space spent during problem solving. Two measures of planning cost have been used in the literature on learning and planning (for example [Minton, 1988, Perez and Etzioni, 1992] namely search time and number of nodes in the search tree. Figure 5 summarizes planning cost criteria. Planning time can be measured as time spent in pure planning, or amortized when planning and learning are interleaved. Planning space is usually measured as working ....

....1992, Kambhampati, 1990, Hammond, 1986] Recycling past successful experience reduces the search effort when solving new similar problems. Note that there is usually a trade off among the amount of knowledge stored, the cost of accessing and reusing it, and the savings on search gained from it [Minton, 1988] . min search tree) working space (e.g. nodes in min (e.g.case library size) long term space min space requirements min planning cost min one time (pure) planning min and learning) amortized time (planning min planning time max (e.g. deriv.anal. recycling of past successful plans Figure 5: A ....

[Article contains additional citation context not shown here]

S. Minton. Learning Effective Search Control Knowledge: An Explanation-based Approach. PhD thesis, Carnegie Mellon University, School of Computer Science, 1988. Also appeared as Technical Report CMU-CS-88-133.

....capability. We think the proposed model formulation method of combining task level analysis and case based reasoning can provide a practical approach for solving ill structured optimization problems. As a limitation in the current status of our research, CABINS suffers from the utility problem [19] since CABINS requires more time for case matching and retrieval with increase in case base size. Although we can define the optimal case base size by monitoring the performance of CABINS for the problems in the domain [22] some knowledge filtering techniques [17] might be useful for improving ....

S. Minton, Learning Effective Search Control Knowledge: An Explanation-Based Approach (Kluwer Academic Publishers, Boston, 1988)

....These notions were verified in experiments reported in [11] To appreciate the results, Figure 4. 1 shows the performance of pure and hybrid refinements in the standard blocks world domain (a non propositional domain) The problems were generated using the random blocks world generator described in [13]. Each data point represents the average over 10 random problems containing a specified number of blocks and where all 10 random problems were solved by the specific instantiation of UCP. UCPFSS, UCP BSS and UCP PS refer to instantiations of UCP using only FSS, BSS and PS refinements respectively. ....

....stands for the state in which block A is on table, block B is on top of block A, block N is on top of block N 1. Problems for the reported experiments are of the type: ffl Go from A ON TOP to N ON TOP state or the other way around. ffl Random blocks world problem as described by Minton [13]. In a problem with N blocks, the goal state can have up to N 2 goal conditions. Goal and initial state are partially specified but the initial information is sufficient to solve the problem. ffl Initial state is a random stack of up to two block height in which the last N 2 blocks are either put ....

S. Minton. Learning effective search control knowledge: An explanation-based approach. PhD thesis, Carnegie-Mellon University, Pittsburgh, PA, 1988.

....state is a collection of random stacks of up to two blocks height in which the last N 2 blocks are either put on the first N 2 blocks or on the table. Goal state is A ON TOP or N ON TOP. 3. Random blocks world problems: A subset of random blocks world problems generated using Minton s algorithm (Minton, 1988). In a problem with N blocks, the goal state can have up to N 2 goal conditions. Some domain dependent pruning tests for blocks world were covered in Section 3. Specifically, we covered pruning test H1 that prevents any block from being moved in consecutive steps, and test H2 which requires that ....

Minton, S. (1988). Learning effective search control knowledge: An explanation-based approach. PhD thesis, Carnegie-Mellon University, Pittsburgh, PA.

....applications that could benefit from effective speedup learning algorithms. 170 T. G. DIETTERICH AND N. S. FLANN 1.1. Explanation Based Learning In the field of machine learning, the best studied speedup learning method is Explanation Based Learning (EBL) as exemplified by the Prodigy (Minton, 1988) and SOAR (Laird, Rosenbloom, Newell, 1986) systems. EBL systems model problem solving as a process of state space search. The problem solver begins in a start state, and by applying operators to the start state and succeeding states, the problem solver seeks to reach a goal state, where the ....

....commit to the region by analyzing the sequence of operators applied in a single experience. This ability to reason with regions has permitted EBL to be applied to problems with infinite state spaces, such as traditional AI planning and scheduling domains, where point based RL would be inapplicable (Minton, 1988). These observations concerning the relationship between EBL and RL suggest that it would be interesting to investigate hybrid algorithms that could perform regionbased backups. These backups would combine the region based reasoning of EBL with the value function approach of RL. The resulting set ....

[Article contains additional citation context not shown here]

Minton, S. (1988). Learning effective search control knowledge: An explanationbased approach. Ph.D. thesis, Carnegie-Mellon University. Technical Report CMU-CS-88-133.

....purposes or goals need not be passed as explicit parameters, but might be stored in the same knowledge base as shared goals. Obtaining expectations about tool performance and distribution of queries is more difficult. This problem might be amenable to machine learning techniques (see, for example, [24]) 9 Conclusion Levesque and Brachman argue that in order to be useful for the most critical applications, general purpose knowledge representation systems should restrict their languages by omitting constructs which require non polynomial (or otherwise unacceptably long) worst case response ....

S. Minton. Learning effective search control knowledge: an explanation-based approach. PhD thesis, Computer Science Department, Carnegie Mellon University, Pittsburgh, Pennsylvania, 1988.

....expositions are readily available (e.g. Jeffrey 1983] The other element of the definition is the agent s fixed constitution or architecture, which sets out the possible states and changes through which learning takes place. We do not have space to elaborate any examples here. See, for example, [Minton 1988], which presents a system that knows enough about its own architecture to allow it to make essentially rational changes in its search strategies on the basis of its search experiences. In addition, precise definitions of many other sorts of architectures are extant, ranging from automata theory ....

....what is rational to do. Of course, mechanization of rational learning can involve calculation of how to change the agent s mental state. This is the natural way to view the long studied hill climbing methods. It also serves as a basis for the bucket brigade algorithm [Holland 1986] and for Minton s [1988] strategy learning system, which collects statistics to estimate expected utilities. Explicit rationality of learning is also reflected in the goal dependent preference order on generalizations employed by [Stepp and Michalski 1986] in the similarity order on analogies employed by [Carbonell ....

Minton, S., 1988. Learning effective search control knowledge: an explanationbased approach, Ph.D. thesis, Computer Science Department, Carnegie Mellon University.

....one shot goal achievement, and it remains to be seen whether it is appropriate for embedding in a longer term context. There is one line of planning research that is not one shot in this sense: work on planning and learning (c.f. the various learning attachments to the Prodigy planning system [11]) This research concerns itself with planners that improve over time; nonetheless, the sense of performance improvement that is relevant is defined in terms of single tasks, rather than in interaction with an environment over time. 1.2.2 Situated Action models Discontents with planning models ....

S. Minton. Learning effective search-control knowledge: An explanationbased approach. Technical Report 133, Carnegie-Mellon University Department of Computer Science, 1988.

....Database Abstractions Although our learning approach is selective, after learning from a large number of queries, the number of the database abstractions could become so large that they degrade the reformulation algorithm s efficiency. This problem is referred to as the utility problem [Minton 88] The utility problem might be alleviated by adopting fast rule match algorithms [Doorenbos et al. 92] such as RETE [Forgy 82] and its more efficient variations. However, if we take the space cost into account, it is still prohibitive to keep a set of the database abstractions that is about the ....

....[Forgy 82] and its more efficient variations. However, if we take the space cost into account, it is still prohibitive to keep a set of the database abstractions that is about the same size or larger than the database just for efficient retrieval. The utility of a rule is defined as follows by [Minton 88] U tility = AverageSaving Theta ApplicationF requency Gamma AverageMatchCost We can measure the utility of learned rules as follows. The AverageMatchCost is proportional to the syntactic length of the rule. The application frequency and average saving can both be computed from statistical ....

Steven Minton. Learning Effective Search Control Knowledge: An Explanation-Based Approach. PhD thesis, Computer Science Department, Carnegie Mellon University, 1988.

....i.e. to automate the acquisition of knowledge that guides the problem solving search process. One approach to learning control knowledge consists of generating explanations for the local decisions made during the search process (DeJong and Mooney, 1986; Laird et al. 1986; Mitchell et al. 1986; Minton, 1988; P erez and Etzioni, 1992; Katukam and Kambhampati, 1994) These explanations become control rules that are used in future situations to prune the search space. These deductive approaches invest a substantial explanation effort to produce provably correct and complete control rules from a single ....

....the search space. The explanation procedure consists of three phases: Labeling the search tree; Credit assignment; and Generation of control rules. The Bounded Explanation module behaves lazily in two aspects: ffl In contrast with eager approaches that learn control knowledge for planning (e.g. (Minton, 1988; Etzioni, 1993) hamlet does not require learning initially correct or complete knowledge. Incremental refinement will be responsible for the correctness of the control knowledge at the end of the learning process. Therefore, there is no need for additional domain axioms. ffl It does not ....

[Article contains additional citation context not shown here]

Minton, S. (1988). Learning Effective Search Control Knowledge: An Explanation-Based Approach.

....systems [ Newell, 1966; Amarel, 1968; Korf, 1980 ] Explicit representation of important information improves the performance of a problem solver. For example, we may improve the efficiency of a problem solving system by encoding useful information about the domain in control rules [ Minton, 1988 ] and an abstraction hierarchy [ Knoblock, 1994 ] On the other hand, explicit representation of irrelevant information decreases efficiency: if we do not mark such information as unimportant for the problem, the system attempts to use it, which takes extra computation and may lead the system to ....

Steven Minton. Learning Effective Search Control Knowledge: An Explanation-Based Approach. PhD thesis, School of Computer Science, Carnegie Mellon University, 1988. Technical Report CMU-CS-88-133.

....deposited and accessible by researchers in machine learning. we address the latter in the context of PRODIGY [Carbonell eta] 1990, Minton eta] 1989b, Ve]oso, 1989] a general purpose complete plan ner that incorporates various learning techniques: explanation based learning (EBL) [Minton, 1988], acquisition of control knowledge through static analysis [Etzioni, 1990] learning by analogy [Veloso, 1991] learning by experimentation [Carbonell and Gil, 1990] learning abstraction hierarchies for effective planning [Knoblock, 1991] and semiautomated knowledge acquisition interfaces ....

....learning time and run time overhead of acquiring and using the new knowledge with increasing complexity. At worst, the overhead should remain a constant fraction of overall problem solving, and at best it should be a diminishing fraction with increased domain size; otherwise the utility problem [Minton, 1988] will prove a serious hindrance. The synergistic combination of multiple learning techniques producing far more performance improvements than individual learning techniques, without paying a correspondingly large overhead cost [Knoblock et al. 1991] Measurements of performance with differ ent ....

[Article contains additional citation context not shown here]

Minton, S. (1988). Learning Effective Search Control Knowledge: An Explanation-Based Approach. PhD thesis, Computer Science Department, Carnegie Mellon University. Available as technical report CMU-CS-88-133.

No context found.

Steven Minton. Learning Effective Search Control Knowledge: An Explanation-Based Approach. PhD thesis, Computer Science Department, Carnegie Mellon University, 1988. 194

No context found.

Steven Minton. Learning Effective Search Control Knowledge: An Explanation-Based Approach. PhD thesis, Computer Science Department, Carnegie Mellon University, 1988.

No context found.

Steven Minton. Learning Effective Search Control Knowledge: An Explanation-Based Approach. PhD thesis, Computer Science Department, Carnegie Mellon University, 1988.

No context found.

Minton, S. (1988). Learning effective search control knowledge: An explanation-based approach. Boston, MA: Kluwer Academic Publishers.

No context found.

Minton, S. 1988. Learning Effective Search Control Knowledge: An Explanation-Based Approach. Boston, MA: Kluwer Academic Publishers.

No context found.

Minton, S. 1988b. Learning Effective Search Control Knowledge: An Explanation-Based Approach. Ph.D. Dissertation, Carnegie-Mellon University, Pittsburgh, PA.

No context found.

Minton, S. 1988a. Learning Effective Search Control Knowledge: An Explanation-Based Approach. Boston, MA: Kluwer Academic Publishers.

No context found.

Minton, S.: 1988, Learning Effective Search Control Knowledge: An Explanation-Based Approach, Boston, MA, Kluwer Academic, Dordrecht.

No context found.

Minton, S., (1988). Learning Effective Search Control Knowledge: An Explanation-based Approach, Doctoral dissertation, Department of Computer Science, Carnegie Mellon Univ.

No context found.

Minton, S. Learning Effective Search Control Knowledge: An explanation-based approach. PhD thesis, Computer Science Department, Carnegie Mellon University, 1988.

No context found.

S. Minton. Learning Effective Search Control Knowledge: An Explanation-Based Approach. PhD thesis, Carnegie-Mellon University, Pittsburgh, PA, 1988.

No context found.

S. Minton. Learning Effective Search Control Knowledge: An Explanation-Based Approach. PhD thesis, Carnegie-Mellon University, Pittsburgh, PA, 1988.

No context found.

S. Minton. Learning Effective Search Control Knowledge: An Explanation-based Approach. Kluwer Academic Publishers, Boston, MA, 1988.

No context found.

Steven Minton. Learning effective search control knowledge: an explanation-based approach. Kluwer, Boston, MA, 1988.

No context found.

Minton, S. 1988a. Learning effective search control knowledge: an explanation-based approach.

First 50 documents  Next 50

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