Kambhampati, S.; Katukam, S.; and Qu, Y. 1996. Failure driven dynamic search control for partial order planners: an explanation based approach. Artificial Intelligence 88(1-2):253--315.  Home/Search   Document Details and Download   Summary   Related Articles   Check

.... guidance [Golding et al. 1987] In prodigy explanation based learning has been applied to learn explicit search control rules [Minton and Carbonell, 1987] Explanation based learning is a technique that has also recently been applied to learning search control rules for a snlp22 like planner [Kambhampati et al. 1996]. The results described there are rather promising as far as the speed up over snlp ( McAllester and Rosenblitt, 1991] cf. section 4.1.2) is concerned. Similarly, Ihrig and Kambhampati, 1997] describe the successful application of explanation based learning to a case based planner. Inductive ....

....if the success rate went below a certain threshold [Minton et al. 1987] Later approaches attempted to approximate the learned search control knowledge to save time [Chase et al. 1989] Wefald and Russell, 1989] have even tried to theoretically define when a search control rule has no benefit. [Kambhampati et al. 1996] have avoided the utility problem by only learning provably correct rules, which are not very many. As far as capability descriptions are concerned, forgetting or approximating search control knowledge means having a less accurate capability description. Considering the advantages of this ....

Subbarao Kambhampati, Suresh Katukam, and Yong Qu. Failure-driven dynamic search control for partial order planners: An explanation-based approach. Artificial Intelligence, 88(1--2):253--315, December 1996.

....called planning by rewriting [1] Under this approach, a partial order planner generates an initial plan, and then a set of rewrite rules are used to transform this plan into a higher quality plan. Unlike the search control rules for partial order planners (such as those learned by UCPOP EBL [6] and PIPP [14] that are defined on the space of partial plans, rewrite rules are defined on the space of complete plans. In addition, it has been argued that plan rewrite rules are easier to state than search control rules, because they do not require any knowledge of the inner workings of the ....

....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 SNLP EBL to learn from planning successes as well as failures. However, these systems only aim to improve planning efficiency and ....

S. Kambhampati, S. Katukam, and Y. Qu. Failure driven dynamic search control for partial order planners. Artificial Intelligence, 88:253--316, 1996.

....generation towards the optimal plan and away from fruitless search, our approach is to generate fast a suboptimal initial plan, and then optimize it, after the fact, by means of the rewriting rules. Explanation Based Learning #EBL# has been used to improve the e#ciency of planning #Minton 1988; Kambhampati, Katukam, Qu 1996#. Our rule generalization algorithm has some elements from EBL, but it compares two complete plans, with the aid of the operator speci#cation, as opposed to problem solving traces. Similarly to EBL search control rules, our learned plan rewriting rules also su#er from the utility problem #Minton ....

Kambhampati, S.; Katukam, S.; and Qu, Y. 1996. Failure driven dynamic search control for partial order planners: an explanation based approach. Arti#cial Intelligence 88#1-2#:253#315.

....driven backtracking. Many different ideas, such backjumping, failure explanation learning and dynamic backtracking can all be concerned as failure driven backtracking . It seems that many quality learning algorithms that learn from failure (c.f. Minton et al. 1989; Bhatnagar Mostow 1994; Kambhampati et al. 1996; Schiex Verfaillie 1993; Dechter, 1990] do analyses the similar way as the type of analysis in the explanation driven backtracking approaches. In order to gain the benefits from the different ideas and approaches related to failure driven backtracking (FDB) and explanation based learning ....

Kambhampati, S., Katukam, S.& Qu, Y., Failure Driven Dynamic Search Control for Partial Order Planner: An Explanation Based Approach, Artificial Intelligence, 88, 1996, 1996

....generation towards the optimal plan and away from fruitless search, our approach is to generate fast a suboptimal initial plan, and then optimize it, after the fact, by means of the rewriting rules. Explanation Based Learning (EBL) has been used to improve the efficiency of planning (Minton 1988; Kambhampati, Katukam, Qu 1996). Our rule generalization algorithm has some elements from EBL, but it compares two complete plans, with the aid of the operator specification, as opposed to problem solving traces. Similarly to EBL search control rules, our learned plan rewriting rules also suffer from the utility problem (Minton ....

Kambhampati, S.; Katukam, S.; and Qu, Y. 1996. Failure driven dynamic search control for partial order planners: an explanation based approach. Artificial Intelligence 88(1-2):253--315.

....solutions. Research in machine learning have long studied the problem of automatically creating e#cient planners by learning domain specific rules or cases to control a general search engine (Minton, 1988; Carbonell, Knoblock, Minton, 1990; Veloso, 1992; Etzioni, 1993; Bhatnagar Mostow, 1994; Kambhampati, Katukam, Qu, 1996; Borrajo Veloso, 1997; Aler, Borrajo, Isasi, 1998; Leckie Zukerman, 1998; etc. However, the successful practical application of machine learning techniques has been limited by at least two factors: First, traditional domain independent planning systems (e.g. PRODIGY, SOAR, NONLIN, UCPOP) ....

Kambhampati, S., Katukam, S., & Qu Y. (1996). Failure driven dynamic search control for partial order planners: An Explanation Based Approach. Artificial Intelligence, 88(1-2), 253--315.

....planners must be able to produce high quality plans, and do so efficiently, if they are to be widely deployed in the real world planning situations. Various approaches have shown that incorporating domain knowledge into domain independent planners can improve both the efficiency of those planners [7, 5, 8] and as well as quality of the plans they produce [14, 6] Traditionally, this knowledge is encoded as search control rules to limit the search for generation of the first viable plan. Recently, Ambite and Knoblock have suggested an alternative approach called planning by rewriting [1] Under this ....

....called planning by rewriting [1] Under this approach, a partial order planner generates an initial plan, and then a set of rewrite rules are used to transform this plan into a higher quality plan. Unlike the search control rules for partial order planners (such as those learned by UCPOP EBL [7] and PIPP [17] that are defined on the space of partial plans, rewrite rules are defined on the space of complete plans. In addition, it has been argued that plan rewrite rules are easier to state than search control rules, because they do not require any knowledge of the inner workings of the ....

[Article contains additional citation context not shown here]

S. Kambhampati, S. Katukam, and Y. Qu. Failure driven dynamic search control for partial order planners. Artificial Intelligence, 88:253--316, 1996.

....to be no harder to write or maintain than any other part of the domain specification. A different (complementary) approach is to automatically generate heuristics. There has been a great deal of work on the problem of heuristic generation, including that on explanation based learning (Minton 1988, Kambhampati et al. 1996), static analysis of operator schemas (Smith 1989, Etzioni 1993) problem abstraction (Knoblock 1994) and operator graph analysis (Smith and Peot 1996) All of this work has aimed at producing explicit search control rules: for example, rules that would state when to prefer one operator ....

Kambhampati, S., Katukam, S., and Qu, Y. (1996). Failure driven dynamic search control for partial order planners: an explanation based approach. Artificial Intelligence 88(1--2), 253-315.

....of mechanisms to guide the threat resolution process of SNLP. Either introducing in SNLP a more elaborate strategy to handle threats (Yang et al. 1992; Peot and Smith, 1993) or, using external control for SNLP such as CBR (Mu noz Avila and Weberskirch, 1996; Ihrig and Kambhampati, 1996) or EBL (Kambhampati et al. 1995). ....

Kambhampati, S., Katukam, S., and Qu, Y. (1995). Failure driven dynamic search control for partial order planners: An explanation-based approach. Artificial Intelligence.

....be analyzed to generate search control rules that explain the failure. When the same situation is encountered the planner will avoid making the choice known to be wrong. We will not provide a detailed description of EBL (in our work we implemented [13] an extension of the mechanism presented in [6]) but, rather, we will illustrate with an example how EBL is used to detect if the failure was caused by an additional goal. Figure 3 sketches the search tree of the situation illustrated in Figure 1 if condition (4) is taken into account. The root of the tree is the leftmost node. The search ....

....initial situation of the new problem. However, the extended subplan cannot be extended to achieve the third goal because of condition (4) From a technical point of view, a 4 The meaning of the term state depends on the particular planning paradigm: for [8] state is a world state whereas for [6] it is a plan state. goal: at(p3,C) at(truck,C) move(truck,C,D) goal: at(truck,C) firstTime(truck,C) goal: firstTime(truck,C) at(truck,C) move(truck,B,C) at(truck,C) move(truck,B,C) at(p3,C) case at(p1,C) at(p2,C) ....

S. Kambhampati, S. Katukam, and Y. Qu. Failure driven dynamic search control for partial order planners: An explanation-based approach. Artificial Intelligence, 88(1-2):253--315, 1996.

....Keane, 1995) Thus, any similarity criterion should measure the adaptation effort of the cases with respect to a new problem. Because the adaptation effort is difficult to determine a priori, learning from previous retrieval episodes has been proposed (Veloso, 1992; Fox and Leake, 1995; Ihrig and Kambhampati, 1995). In Robbie (Fox and Leake, 1995) introspective reasoning (Leake et al. 1995) is used to determine the features that should be considered during retrieval. The validity of pre defined assertions related to indexing criteria is tested after each retrieval episode. If a failure occurs, the ....

....assertions related to indexing criteria is tested after each retrieval episode. If a failure occurs, the criteria is restated resulting in a refinement of the index. In the context of domain independent planning, EBL (Minton, 1988) has been used in a system called derSNLP EBL (Ihrig and Kambhampati, 1995) to explain retrieval failures. An EBL rule that indicates combinations of features causing a failure is constructed. The rule is used as a filter to avoid selecting the case when a given problem matches the indicated combination of features. Another approach toward learning from retrieval ....

[Article contains additional citation context not shown here]

Kambhampati, S., Katukam, S., and Qu, Y. (1995). Failure driven dynamic search control for partial order planners: An explanation-based approach. Artificial Intelligence.

No context found.

S. Kambhampati, S. Katukam, and Y. Qu. Failure driven dynamic search control for partial order planners: An explanation-based approach. Artificial Intelligence, 88, 1996.

No context found.

S. Kambhampati, S. Katukam and Y. Qu. Failure driven dynamic search control for partial order planners: An explanationbased approach. Artificial Intelligence, Fall 1996. (To appear).

No context found.

S. Kambhampati, S. Katukam, and Y. Qu. Failure driven dynamic search control for partial order planners: An explanation-based approach. Artificial Intelligence,1996. To Appear.

....and dynamic backtracking are all concerned with the general notion of intelligent backtracking. Complicating the picture further is the fact that many speedup learning algorithms that learn from failure (c.f. Minton, Carbonell, Knoblock, Kuokka, Etzioni, Gil, 1989; Bhatnagar Mostow, 1994; Kambhampati, Katukam, Qu, 1996; Dechter, 1990; Schiex Verfaillie, 1993) do analyses that are quite close to the type of analysis done in the intelligent backtracking algorithms. Although this similarity has sometimes been noted in earlier literature (c.f. Dechter, 1990) a thorough analysis has been impeded by the many ....

....mechanisms 2 can outperform non systematic searchers such as GSAT and WALKSAT (Selman et al. 1992) on several hard real and artificial satisfiability instances. Similarly, within the planning and problem solving communities, EBL approaches are finding continued uses in learning search control (Kambhampati et al. 1996), case based planning (Ihrig Kambhampati, 1996; Munoz Avila Weberskirsch, 1996) plan quality control (Estlin Mooney, 1996) Moreover, recent work in planning has amply emphasized the role of constraint satisfaction in plan synthesis (Kautz Selman, 1996; Kambhampati Yang, 1996; Joslin ....

[Article contains additional citation context not shown here]

Kambhampati, S., Katukam, S., & Qu, Y. (1996). Failure driven dynamic search control for partial order planners: An explanation-based approach. Artificial Intelligence, 88.

.... very large search spaces even for simple problems [7, 2] The usual solution to this problem is to control the planner s search with the help of search control knowledge acquired from domain experts (e.g. task reduction schemas) or through learning techniques (e.g. explanation based learning [15], case based planning [6] In this paper, we will consider a more direct solution to the search space explosion problem that of handling sets of plans without splitting them into the search space. At first glance, this seems to involve a mere exchange of complexity from search space size to ....

S. Kambhampati, S. Katukam, Y. Qu. Failure Driven Dynamic Search Control for Partial Order Planners: An explanation-based approach Artificial Intelligence (To appear in Fall 1996)

....practical planners use depthfirst rather than best first search strategies for efficiency purposes. Looping can significantly affect the efficiency of depth first search regimes. The final, and often overlooked, need for pruning strategies has got to do with the importance of failures in learning [10]. Most speedup learning strategies to improve planning performance learn from the failures encountered in plan generation. Existence of a variety of pruning techniques provides a rich opportunity for the learner to learn from the pruned branches. Although best first search strategies may be able ....

....In the presence of learning strategies, the cost of pruning techniques also tends to be less of a concern. In particular, it is possible to use the pruning techniques strategically by combining them with a depth limited search, and applying them only to the plans that cross the depth limits (see [10] for a demonstration of the effectiveness of this approach) Despite their importance, very little work has been done towards formulation and evaluation of pruning techniques for plan space planning. Most existing plan space planners prune a plan when the constraints on the partial plan are ....

[Article contains additional citation context not shown here]

S. Kambhampati, S. Katukam and Y. Qu. Failure Driven Dynamic Search Control for Partial Order Planners: An Explanation Based Approach. Artificial Intelligence, To appear. (Available from http://rakaposhi.eas.asu.edu:8001/yochan.html).

....0 = d Gamma1 (E 1 ) we can learn a rule which recommends rejection of the decision d whenever E 0 is present in the current node. 3 3 Explanation basedlearning normally also involves a generalization step, where the failure explanation is generalized by replacing constants with variables [9] . Although such generalization can be very important in supporting inter problem transfer, addition of generalization steps does not have a crucial impact on the analysis given in this paper. See [9] for a comprehensive discussion of the issues involved in explanation generalization. x= A = w ....

....step, where the failure explanation is generalized by replacing constants with variables [9] Although such generalization can be very important in supporting inter problem transfer, addition of generalization steps does not have a crucial impact on the analysis given in this paper. See [9] for a comprehensive discussion of the issues involved in explanation generalization. x= A = w = E y=B = u = D u = C = A v = D = l = B x ,y,u,v: A, B C, D, E w : D E : A B Domains: Constraints: Problem Spec. 0 B x A y B u C N1: x = A N2: x = A y = N3: x=A , y ....

[Article contains additional citation context not shown here]

S. Kambhampati, S. Katukam and Y. Qu. Failure driven dynamic search control for partial order planners: An explanationbased approach. Artificial Intelligence, Fall 1996. (To appear).

....by incorporating explanationbased learning (ebl) techniques for detecting and explaining analytical failures in the planner s search space. These include methods for forming explanations of search path failures and regressing these explanations through the planning decisions in the failing paths (Kambhampati, Katukam, Qu 1996). Here we employ these techniques to construct reasons for case failure, which are then used to annotate the failing cases to constrain their future retrieval. Furthermore, each failure results in the storage of a new case which repairs the failure. dersnlp ebl normally stores plan derivations ....

....that experiences a failure and toward the case that repairs the failure. We are now in a position to describe how the planner learns the reasons underlying a case failure. Specifically, we use EBL techniques to accomplish this learning. In the next section, we show how the techniques developed in (Kambhampati, Katukam, Qu 1996) are employed to construct these reasons. Case Failure Explanation: C = fh(AT OB OB1 l d ) tGi h(AT OB OB2 l d ) tGig E = fhtI ; AT OB OB2 l d )i htI ; INSIDE PL OB2 PL )i htI ; AT OB OB2 l 1 )i htI ; AT OB OB2 l p )ig Figure 4: An example of a case failure reason Learning from ....

[Article contains additional citation context not shown here]

Kambhampati, S.; Katukam, S.; and Qu, Y. 1996. Failure driven dynamic search control for partial order planners: An explanation-based approach. Artificial Intelligence. To Appear.

No context found.

Kambhampati, S.; Katukam, S.; and Qu, Y. 1996. Failure driven dynamic search control for partial order planners: an explanation based approach. Artificial Intelligence 88(1-2):253--315.

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