Laird, J., Rosenbloom, P. & Newell, A. Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies, Boston, Mass: Kluwer Academic Publishers. Lakoff, G. (1986). Connectionism and congitive linguistics. Seminar delivered at Princeton University, December 8, 1986.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....They may simply not be smart enough to know what to do when a limited stock of methods fails to apply. But this needn t be a principled limitation of Classical architectures: There is, to our knowledge, no reason to believe that something like Newell s (1969) hierarchy of weak methods or Laird, Rosenberg and Newell s (1986) universal subgoaling, i s n principle incapable of dealing with the problem of graceful degradation. Nor, to our knowledge, has any argument yet been offered that Connectionist architectures are in principle capable of dealing with it. In fact current Connectionist models are every bit as ....

Laird, J., Rosenbloom, P. & Newell, A. Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies, Boston, Mass: Kluwer Academic Publishers. Lakoff, G. (1986). Connectionism and congitive linguistics. Seminar delivered at Princeton University, December 8, 1986.

.... Faced with brittleness problems in the Generic Task architecture (Chandrasekaran, 1986) investigators at Ohio State University are developing to an approach in which generic tasks are implemented in SOAR (Johnson et al. 1990) Each generic task is represented within a SOAR problem space (Laird et al. 1986). As such, the inputs to the generic task are modeled as the initial state of the problem space; the desired output of the generic task is modeled as the goal state; and the problem solving method is modeled using operators available within the problem space. The Ohio State group hopes that ....

Laird, J., Rosenbloom, P., Newell, A. (1986). Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies. Boston: Kluwer Academic.

....(possibly wrong) heuristic response, an agent can revisit previous decision situations when it has the time and use RMM to determine what the rational response should have been. By storing this as a rule of behavior that can be recalled when appropriate in the future (see related work on chunking [47]) RMM can provide the basis for the accrual of rational heuristics [58] Another important direction, and an application area, of RMM is studying rational communicative behavior among agents involved in interactions. It turns out that our framework allows the agents to also compute the expected ....

John Laird, Paul Rosenbloom, and Allen Newell. Universal Subgoaling and Chunking: The automatic generation and learning of goal hierarchies. Kluwer Academic Publishers, 1986.

....in solving a problem instance or class of instances. For problems which require a solution to satisfy a set of constraints, the best heuristic is typically one which is expected to allow the problem solver to most efficiently converge on a solution. For example, the approach exemplified by SOAR [ Laird et al. 1986 ] uses traces of past problemsolving efforts to refine heuristics in order to speed up the search for a solution. However, other criteria besides problem solving efficiency for heuristic selection are possible. For scheduling and other constrained optimization problems, quality of solution may ....

John E. Laird, Paul S. Rosenbloom and Allan Newell. Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Heuristics. Norwell, MA: Klewer Academic Publishers, 1986.

....means ends tasks, cognition must have a task goal stack available. This assumption was embedded in the problem behavior graph and the General Problem Solver (Newell Simon, 1972) as representations of human behavior, was later incorporated in cognitive architectures (Anderson, 1983; Laird, 1984; Laird, Rosenbloom, Newell, 1986), and lives on today in many forms. The goal stack is a feature of task analysis tools like GOMS (John Kieras, 1996) and remains a core mechanism in cognitive architectures like ACT R (Anderson Lebire, 1998) and Soar (Lehman et al. 1998; Newell, 1990; Newell, Yost, Laird, Rosenbloom, ....

Laird, J. E., Rosenbloom, P. S., & Newell, A. (1986). Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies. Boston: Kluwer Academic Publishers.

....in solving a problem instance or class of instances. For problems which require a solution to satisfy a set of constraints, the best heuristic is typically one which is expected to allow the problem solver to most efficiently converge on a solution. For example, the approach exemplified by SOAR [ Laird, et al. 1986 ] uses traces of past problem solving efforts to refine heuristics in order to speed up the search for a solution. However, other criteria besides problem solving efficiency for heuristic selection are possible. For scheduling and other constrained optimization problems, quality of solution may be ....

John E. Laird, Paul S. Rosenbloom and Allan Newell. Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Heuristics. Norwell, MA: Klewer Academic Publishers.

.... written in a shell that uses operators corresponding to Prolog s cut or not( Delta) as well as any system that returns only the first answer found; this class of shells includes TestBench 1 and other fault hierarchy systems, prioritized default theories [6, 29] most production systems [33, 20], as well as Prolog [8] The goal of a theory revision process is to improve the accuracy of the reasoning system on its performance task of answering queries. Section 2 first defines this objective more precisely: as identifying the revision (i.e. sequence of transformations ) that produces a ....

John E. Laird, Paul S. Rosenbloom, and Allan Newell. Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies. Kluwer Academic Press, Hingham, MA, 1986.

....in response to careful, automatic, analysis of past problem solving attempts. Generative approaches consider not only the structure of the domain, but also structures that arise from the problem solver interacting with specific problems from the domain. This approach is exemplified by SOAR [Laird86] and the PRODIGY EBL system of Steve Minton [Minton88] These techniques analyze 5 past problem solving traces and conjectures heuristic control rules in response to specific problem solving inefficiencies. Such approaches can effectively exploit the idiosyncratic structure of a domain through ....

J. E. Laird, P. S. Rosenbloom and A. Newell, Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies, Kluwer Academic Publishers, Hingham, MA, 1986.

....solving situations. A general algorithm might learn the idiosyncrasies of a domainthrough its problem solving experiences and automatically transform itself into an effective problem solver. In fact, machine learning techniques have successfully demonstrated this capacity in limited contexts [Fikes71, Laird86, Minton88]. Nonetheless, adaptive problem solving is still far from realization in any general sense. The principal impediment to adaptive problem solving is characterizing when an automatically hypothesized transformation actually results in improved problem solving performance. Steve Minton introduced to ....

....structure to transformations that can be exploited. In most learning techniques a composite transformation consists of many individual atomic transformations (later we refer to atomic modifications simply as transformations) For example, SOAR constructs a new problem solver from individual chunks [Laird86]. PRODIGY EBL builds a learned control strategy from individual control rules [Minton88] Two different composite transformations may share many of the same individual components. Instead of making a global decision among all possible composite transformations, an incremental learning system ....

J. E. Laird, P. S. Rosenbloom and A. Newell, Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies, Kluwer Academic Publishers, Hingham, MA, 1986.

....previously solved problems, usually available as a set of cases, to come to abstract solutions. The use of experience has already proven useful in various approaches to speedup learning such as explanation based learning (Mitchell, Keller, Kedar Cabelli, 1986; DeJong Mooney, 1986; Rosenbloom Laird, 1986; Minton, 1988; Minton, Carbonell, Knoblock, Kuokka, Etzioni, Gil, 1989; Shavlik O Rorke, 1993; Etzioni, 1993; Minton Zweben, 1993; Langley Allen, 1993; Kambhampati Kedar, 1994) and analogical or case based reasoning (Carbonell, 1986; Kambhampati Hendler, 1992; Veloso Carbonell, ....

....for successful solutions already known by the system. In explanation based approaches, these explanations mostly cover the whole problem solving process (Fikes, Hart, Nilsson, 1972; Mooney, 1988; Kambhampati Kedar, 1994) but can also relate to to problem solving chunks (Rosenbloom Laird, 1986; Laird, Rosenbloom, Newell, 1986) of some smaller size or even to single decisions within the problem solving process (Minton, 1988; Minton et al. 1989) Explanation based approaches generalize the constructed explanations during learning by extensive use of the available domain knowledge and ....

[Article contains additional citation context not shown here]

Laird, J., Rosenbloom, P., & Newell, A. (1986). Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies. Kluwer Academic Publishers, Norwell, MA.

....solve problems. This can be viewed as a transformational process where a series of transformations are applied to the original problem solver (see [Gratch90, Greiner92] Aplannermaybetransformed by the addition of control knowledge. Different forms of control knowledge include macro operators [Braverman88a, Laird86, Markovitch89], control rules [Etzioni90a, Minton88, Mitchell83] and static board evaluation functions [Utgoff91] Alternatively, a learning system may modify the domain theory. Such transformations could be truth (accuracy) preserving (as in conjunct reordering or deletion of redundancy [Smith85] or ....

....it restricts the class of transformations which are actively considered. A common approach is event driven learning. Under this strategy, transformations are only proposed in response to planning events such as success or failure which are observed in the course of problem solving. ForexampleSOAR [Laird86] only learns new chunks in response to planning impasses. STRIPS [Fikes72] only acquires new macro operators from successful plans. Even given a single planning event, there may be many transformations which could be proposed. This complexity is frequently handled by a heuristic selection ....

[Article contains additional citation context not shown here]

J. E. Laird, P. S. Rosenbloom and A. Newell, Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies, Kluwer Academic Publishers, Hingham, MA, 1986.

....system s strengths and weaknesses. 1 INTRODUCTION In machine learning there is considerable interest in techniqueswhichimproveplanning ability. Investigation in this area has identified a wide array of techniques including macro operators [DeJong86, Fikes72, Mitchell86, Segre88] chunks [Laird86], and control rules [Minton88, Mitchell83 ] With these techniques comes a growing battery of successful demonstrations in domains ranging from 8 puzzle to space shuttle payload processing. Unfortunately, the formal properties of these approaches are not wellunderstood. This is highlighted by ....

....does not hold, this simplification can sacrifice minimal adequacy. For example, we document how negative interactions between control rules cause PRODIGY EBL [Minton88 ] to to acquire harmful control strategies. Many learning systems do not explicitly consider interactions (including SOAR [Laird86], STATIC [Etzioni90b] PRODIGY EBL, RECEBG [Letovsky90] IMEX [braverman88 ] and PEBL [Eskey90] Ma and Wilkins illustrate a similar situation for knowledge base revision systems [Wilkins89] Systems which do not adopt this simplification include PALO [greiner92] COMPOSER [Gratch92b] and ....

J. E. Laird, P. S. Rosenbloom and A. Newell, Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies, Kluwer Academic Publishers, Hingham, MA, 1986.

....In small domains with well defined parameters (e.g. 8 puzzle game) the system will perform less efficiently over time than other learning systems. The decrease in efficiency will occur due to the presence of compiled formats learned from previous experience like macro operators[Korf83] chunks [Laird86], or concepts [DeJong86] in the other systems. The number of complied formats tend not to increase due to the simplicity of the problem and its static nature. These systems will not get cluttered with compiled formats, hence their performance will not degrade. In SEIF no such form of compiled ....

Laird, John E., Rosenbloom, P. S., Newell, Alan, (1986). "Universal Subgoaling and Chunking: The automatic generation and learning of goal hierarchies." In Machine Learning Journal, Hingham, MA, Kluwer Academic Publishers.

....In small domains with well defined parameters (e.g. 8 puzzle game) the system will perform less efficiently over time than other learning systems. The decrease in efficiency will occur due to the presence of compiled formats learned from previous experience like macro operators[Korf83] chunks [Laird86], or concepts [DeJong86] in the other systems. The number of complied formats tend not to increase due to the simplicity of the problem and its static nature. These systems will not get cluttered with compiled formats, hence their performance will not degrade. In SEIF no such form of compiled ....

Laird, John E., Rosenbloom, P. S., Newell, Alan, (1986). "Universal Subgoaling and Chunking: The automatic generation and learning of goal hierarchies." In Machine Learning Journal, Hingham, MA, Kluwer Academic Publishers.

....q) Gamma c( Theta 0 ; q) If there is Theta 0 2 T [ Theta k ] s.t. Delta[ Theta 0 ; Theta k ] ffl 2 Then Let Theta k 1 Theta 0 Else Return[ Theta k ] End For End palo Figure 1: palo Algorithm Many speed up learning systems fit into this framework; cf. MKKC86, DM86] LRN86] Most of these systems, however, climb to a new performance element (by incorporating a new macro, or an additional control heuristic, etc. after observing a single query; in each case, forming a new performance system that would work better on that specific query . Unfortunately, the resulting ....

John E. Laird, Paul S. Rosenbloom, and Allan Newell. Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies. Kluwer Academic Press, 1986.

.... written in a shell that uses operators corresponding to Prolog s cut or not( Delta) as well as any system that returns only the first answer found; this class of shells includes TestBench 1 and other fault hierarchy systems, prioritized default theories [6, 28] most production systems [32, 20], as well as Prolog [8] The goal of a theory revision process is to improve the accuracy of the reasoning system on its performance task of answering queries. Section 2 first defines this objective more precisely: as identifying the revision (i.e. sequence of transformations ) that produces a ....

John E. Laird, Paul S. Rosenbloom, and Allan Newell. Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies. Kluwer Academic Press, Hingham, MA, 1986.

.... of a performance system on its performance task, discussed in [28] and elsewhere, is the basis for learning systems that range from standard decision tree learners like c4.5 [26] and cart [3] that seek the decision tree whose expected accuracy is maximal, through speed up learning systems [25] [21] that seek a set of macros that yield an optimally efficient program, to neural net learners [27] 24] 14] that seek a setting of the weights that produces an optimal classifier, etc. Moreover, LearnSF qualifies as a wrapper learning system [17] 4] as it views its performance elements ....

J. E. Laird, P. S. Rosenbloom, and A. Newell, Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies. Kluwer Academic Press, Hingham, MA, 1986.

....(possibly wrong) heuristic response, an agent can revisit previous decision situations when it has the time and use RMM to determine what the rational response should have been. By storing this as a rule of behavior that can be recalled when appropriate in the future (see related work on chunking [44]) RMM can provide the basis for the accrual of rational heuristics. Further, multiple research issues arise as to how decision theoretic reasoning about actions of other agents can be integrated with other forms of reasoning, for example with nested models of the other agents deduction (see, for ....

John Laird, Paul Rosenbloom, and Allen Newell. Universal Subgoaling and Chunking: The automatic generation and learning of goal hierarchies. Kluwer Academic Publishers, 1986.

....values in the case of tree like decision structures, and proved that (unless RP = NP ) there can be no efficient algorithm for this task for general structures. Our approach also resembles the work on speed up learning (including both explanationbased learning [15, 5, 14] and chunking [13]) as it uses previous solutions to suggest a way of improving the speed of a performance system. Most speed up learning systems, however, use only a single example to suggest an improvement; we extend those works by showing how to use a set of samples and by describing, furthermore, the exact ....

John E. Laird, Paul S. Rosenbloom, and Allan Newell. Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies. Kluwer Academic Press, Hingham, MA, 1986.

....implemented in the XAPS2 production system architecture; that is, the same architecture upon which Soar 1 was based. Over the next year the model was extended and made task independent, and became a central part of Rosenbloom s thesis [ Rosenbloom, Figure 5: Structure of the Weak Methods 1983; Laird et al. 1986b ] This extended, task independent chunking model was implemented as part of a new productions system architecture that was designed specifically to support it XAPS3. During the summer and fall of 1983, we discussed how to incorporate chunking into Soar, and finally in January of 1984, we ....

....where it is directly available. Although this type of learning affects performance at the problem space and symbol levels, it has been argued that along with explanation based learning (EBL) Mitchell et al. 1986; DeJong and Mooney, 1986 ] to which chunking is closely related [ Rosenbloom and Laird, 1986 ] it does not produce learning at the knowledge level [ Dietterich, 1986 ] However, the argument is based on the mistaken notion that, at the knowledge level, a system knows everything that it can derive at the symbolic architecture level from its symbol structures and the processes that ....

[Article contains additional citation context not shown here]

J. E. Laird, P. S. Rosenbloom, and A. Newell. Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies. Kluwer Academic Publishers, Hingham, MA, 1986.

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