The chunking of skill and knowledge

Abstract

This article describes recent work that utilizes the concept of chunking as the basis for an integrated model of the acquisition of both skill and knowledge. We look at results in the areas of practice (skill) and verbal learning (knowledge). The approach is based on viewing task performance as a problem solving process and chunking as a learning process that stores away information about the results of problem solving. In practice tasks, chunks acquired during the solution of one problem can be used during later problems to speed up the system's performance. This chunking process produces the same type of power-law practice curves that appear so ubiquitously in human practice. In verbal learning tasks, chunks acquired during training are used at test time to determine how to respond. This psychological model is a manifestation of a set of processes that provide the basis of a general architecture. Such an architecture is not only interesting in its own right, but provides support for the more narrowly based psychological phenomena. The concept of chunking has played a major theoretical role in cognitive psychology ever since Miller's classic paper (Miller, 1956). Through much of this history it has been used primarily in models of memory organization. According to this view, chunking is a process of creating symbols (chunks) which represent the combination of several other symbols. In this long and productive tradition, chunking has been used to model a wide variety of memory phenomena (Miller, 1956; DeGroot, 1965; Bower & Winzenz, 1969; Johnson, 1972; Chase & Simon, 1973; Chase & Ericsson, 1981). In recent years, chunking has also been proposed as the basis for a model of human This research was sponsored by the Defense Advanced Research Projects Agency (DOD) under contract N00039-86-C-0133 and by the Sloan Foundation. Computer facilities were partially provided by NIH grant RR-00785 to Sumex-Aim. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Defense Advanced Research Projects Agency, the US Government, the Sloan Foundation, or the National Institutes of Health. Knowledge Systems Laboratory, Departments of Computer Science and Psychology, Stanford University, 701 Welch Road (Bldg. C), Palo Alto, CA 94304 (After September 1987, University of Southern California, Information Sciences Institute, 4676 Admiralty Way, Marina del Rey, CA 90292) Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI 48109 A Department of Computer Science, Carnegie-Mellon University, Pittsburgh, PA 15213 practice (Newell & Rosenbloom, 1981; Rosenbloom, 1983; Rosenbloom & Newell, 1987a). According to this view, chunking is a process that acquires new productions (chunks) by the caching of goal-based performance. This version of chunking was successfully able to model the ubiquitous power law shape of human practice curves. In building on these results, chunking has been reimplemented as a learning mechanism within an artificial intelligence problem solver called Soar (Laird, 1983; Laird, 1986; Laird, Newell, & Rosenbloom, 1987). This combination of Soar and chunking has provided a learning architecture with the flexibility and power needed to demonstrate learning behaviors that go considerably beyond simple practice effects (Steier et al, 1987). The results have been good enough for chunking to be proposed as a general learning mechanism for artificially-intelligent systems (Laird, Rosenbloom, & Newell, 1984; Laird, Rosenbloom, & Newell, 1986). The generality of the combination of chunking and Soar as an AI learning system raises the possibility that this combination can be transferred back into psychology to provide the basis for a general model of human learning. In this article we report on a preliminary examination of this possibility that involves the application of Soar to two major domains of human learning: skill acquisition and knowledge acquisition. Within skill acquisition we return to the subdomain of practice, asking whether Soar can replicate the earlier results by producing power law practice curves. Within knowledge acquisition we look within the subdomain of verbal learning at simple recognition, recall, and cued recall tasks. In this preliminary investigation we are not yet looking for close matches to experimental data. Instead, we focus on the nature of the processing that allows these tasks to be performed by chunking. In the next two sections we give brief descriptions of Soar and chunking. This background material is followed by sections on skill acquisition and knowledge acquisition in Soar, and then by a section containing concluding remarks. 1. Soar Soar is based on formulating all symbolic goal-oriented processing as search in problem spaces (Newell, 1980). The problem space determines the set of states and operators that can be used during the processing to attain a goal. The states represent situations. There is an initial state, representing the initial situation, and a set of desired states that represent the goal. An operator, when applied to a state in the problem space, yields another state in the problem space. The goal is achieved when a desired state is reached as the result of a sequence of operator applications starting from the initial state. Each goal defines a problem solving context ("context" for short). A context is a data structure in Soar's working memory — a short-term declarative memory — that contains, in addition to a goal, roles for a problem space, a state, and an operator. Problem solving for a goal is driven by the acts of selecting problem spaces, states, and operators for the appropriate roles in the context. Each deliberate act of the Soar architecture — a selection of a problem space, a state or an operator — is accomplished via a two-phase decision cycle. First, during the elaboration phase, the description of the current situation (that is, the contents of working memory) is elaborated with relevant information from Soar's production memory — a long-term procedural memory. The elaboration phase proceeds in a sequence of synchronous cycles. During each cycle of the elaboration phase, all of the productions in the production memory are matched against working memory, and then all of the resulting production instantiations are executed. The net effect of these production firings is to add information to the working memory. New objects are created, new knowledge is added about existing objects, and preferences are generated. There is a fixed language of preferences that is used to describe the acceptability and desirability of the alternatives being considered for selection. By using different preferences, it is possible to assert that a particular problem space, state, or operator is acceptable (should be considered for selection), rejected (should not be considered for selection), better than another alternative, and so on. When the elaboration phase reaches quiescence — that is, no more productions can fire — the second phase of the decision cycle, the decision procedure, is entered. The decision procedure is a fixed body of code that interprets the preferences in working memory according to their fixed semantics. If the preferences uniquely specify an object to be selected for a role in a context, then a decision can be made, and the specified object becomes the current value of the role. The decision cycle then repeats, starting with another elaboration phase. If, when the elaboration phase reaches quiescence, the preferences in working memory are either incomplete or inconsistent, an impasse occurs in problem solving because the system does not know how to proceed. When an impasse occurs, a subgoal with an associated problem solving context is automatically generated for the task of resolving the impasse. The impasses, and thus their subgoals, vary from problems of selection (of problem spaces, states, and operators) to problems of generation (e.g., operator application). Given a subgoal, Soar can bring its full problem solving capability and knowledge to bear on resolving the impasse that caused the subgoal. When impasses occur within impasses, then subgoals occur within subgoals, and a goal hierarchy results (which therefore defines a hierarchy of contexts). The top goal in the hierarchy is a task goal; such as, to recognize an item. The subgoals below it are all generated as the result of impasses in problem solving. A subgoal terminates when its impasse (or some higher impasse) is resolved. 2. Chunking The traditional view of chunks is that they are symbols representing the combination of several other symbols (Miller, 1956). Newell and Rosenbloom (1981) turned this concept into a model of practice by describing how performance could be improved by the acquisition of chunks that represent patterns of objects in the task environment. This model was instantiated in a production systemarchitecture, first in a domainspecific form (Rosenbloom & Newell, 1987b), and then in a domain-independent form (Rosenbloom, 1983; Rosenbloom & Newell, 1987a). A modified version of the domainindependent chunking mechanism was then implemented as part of the Soar architecture (Laird, Rosenbloom, & Newell, 1986). In its Soar implementation, chunking is a learning mechanism that acquires new productions which summarize the processing that leads to results of subgoals. The actions of the new productions are based on the results of the subgoal. The conditions are based on those aspects of the pre-goal situation that were relevant to the determination of the results. Relevance is determined by using the traces of the productions that fired during the subgoal. Starting from the production trace that generated the subgoal's result, those production traces that generated the workingmemory elements in the conditions of the trace are found, and then the traces that generated their condition elements are found, and so on until elements are reached that existed prior to the subgoal. Productions that only generate preferences do not participate in this backtracking process — preferences only affect the efficiency with which a goal is achieved, and not the correctness of the goal's results. An example of this chunking process is shown schematically in Figure 2-1. The circled letters are objects in working memory. The two striped vertical bars mark the beginning and ending of the subgoal. The objects to the left of the first bar (A, B, C, D, E, and F) exist prior to the creation of the subgoal. The objects between the two bars (G, H, and I) are internal to the subgoal. The objects to the right of the second bar (J) are results of the subgoal. P i , P2, P3, and P4 are production traces; for example, production trace Pi records the fact that a production fired which examined objects A and B and generated object G. The highlighted production traces are those that are involved in the backtracing process. Chunking in this figure begins by making the result object (J) the basis for the re 2-1: Schematic view of the chunking process in Soar. action of the chunk. The condition-finding process then begins with object J, and determines which production trace produced it — trace P4. It then determines that the conditions of trace P4 (objects H and I) are generated by traces P2 and P3, respectively. The condition elements of traces P2 and P3 (objects C, D, E, and F) existed prior to the subgoal, so they form the basis for the conditions of the chunk. The resulting chunk is C A D A E A F — > J. (1) Once a chunk has been learned, the new production will fire during the elaboration phase in relevantly similar situations in the future, directly producing the required information. No impasse will occur, and problem solving will proceed smoothly. Chunking is thus a form of goal-based caching which avoids redundant future effort by directly producing a result that once required problem solving to determine. It bears a family resemblance to other learning mechanisms which move the system along the store-versus-compute trade-off, such as production composition (Lewis, 1978; Neves & Anderson, 1981; Anderson, 1982; Anderson, 1983), memo functions '(Michie, 1968), macro-operators (Fikes, Hart, & Nilsson, 1972; Korf, 1985), and explanation-based generalization (Mitchell, Keller, & Kedar-Cabelli, 1986). In the pre-Soar implementation of chunking, chunks were learned bottom-up in the goal hierarchy, reflecting the accepted view of human chunking. On any trial, only the terminal subgoals — the subgoals that did not themselves have subgoals — were chunked. Higher-level subgoals could be chunked only after all of their subgoals had been chunked. This bottom-up approach was critical in producing power law practice curves. However, chunking has been implemented in Soar with a settable option. Chunking can proceed for only the terminal subgoals (bottom-up chunking) or for all of the goals in the hierarchy (all-goals chunking). Given a sufficient number of trials, the results of the two approaches should be essentially indistinguishable. However, because all-goals chunking yields faster learning, most research on learning in Soar has employed it. In this article, all results except for those specifically intended to model practice curves were generated with all-goals chunking. Practice curves were generated with bottom-up chunking. 3, Skill Acquisition There is a ubiquitous regularity in human practice — the power law of practice — that states that the time to perform a task (T) decreases as a power-law function of the number of times the task has been performed (that is, the number of trials, N): While the power law of practice was originally recognized in the domain of motor skills (Snoddy, 1926), it has recently become clear that it holds over a much wider range of human tasks — possibly extending to the full range of human performance (Newell & Rosenbloom, 1981). The driving force behind the initial development of chunking as a learning mechanism was a desire to develop a model of practice that would produce power law practice curves. The experimental task used during these early investigations was a 1023-choice reaction-time task (Seibel, 1963). This is a perceptual-motor task in which there is a stimulus display of ten lights, arranged horizontally, and a response apparatus of ten buttons, arranged horizontally in such a way that each finger rests on one of them. The stimulus and response environments are set up so that there is a highly compatible one-one correspondence between the lights and buttons, each light directly above a button. On each trial of the experiment, some of the lights are on and some are off. The subject's task is to respond as quickly as possible by pressing the buttons corresponding to the lights that are on. The performance algorithm used in modeling this task is based on a top-down divideand-conquer algorithm. The top goal is to process a horizontal region of the display, containing a pattern of on and off lights. This goal is recursively decomposed into subgoais to process smaller and smaller regions until patterns are reached which can be directly converted into corresponding patterns of button presses. Chunking works in a bottom-up fashion in this goal hierarchy. Productions are first learned for the bottommost level of subgoals. These chunks relate small patterns of lights to their corresponding patterns of button presses. In later trials, these chunks can be used in place of problem solving in the low-level subgoals, speeding up task performance. This also allows chunking to proceed up the goal hierarchy, acquiring chunks which relate increasingly larger patterns of lights to patterns of button presses. A simple analysis of this learning process suggests exponential speed-ups with

2 Figures and Tables

Cite this paper

@inproceedings{Rosenbloom2016TheCO, title={The chunking of skill and knowledge}, author={Paul S. Rosenbloom and John E. Laird and Allen Newell}, year={2016} }