Fragments of the Theory

Let me provide a few quick illustrations of the theory. These will be like strobe-light exposures—a fragment here, a flash there. Still, I hope they can bring home two critical points. First, Soar is a theory, in the same mold as theories in the other sciences, a collection of mechanisms that combine togethet to predict and explain empirical phenomena. The predictions come from the theory, not the theorist. Second, as a unified theory of cognition, Soar has a wide scope, both in types of behavior covered and in terms of time scale. Though never as great as wishes would have it, Soar can still stand for the possibility that unified theories of cognition might be in the offing. Let us begin with immediate reactive behavior, which occurs at a time scale of about 1 sec., and work up the time scale of human action.

Stimulus-Response Compatibility

Stimulus-response compatibility is a phenomenon known to everyone, though perhaps not by that name. Anyone who has arrived at an elevator to find the Up button located physically below the Down button would recognize the phenomena. The Up button should map into the direction of travel—up on top. This human sense of should , in fact translates into longer times to hit the button and greater chances to hit the wrong button. Stimulus-response compatibility effects are everywhere. Figure 9.3 shows another example, perhaps less obvious. A person at a computer editor wants to delete some word. The editor uses abbrevations, in this case dropping the vowels to get dlt . Thus, the

Fig. 9.3.
SRC example: Recall command abbreviation

person needs to get from delete to dlt to command the editor appropriately. Stimulus-response compatibility occurs here. On the more compatible side, the designer of the editor might have chosen delete itself, although it would have required more typing. On the less compatible side, the designer might have chosen gro , thinking of get rid of .

Figure 9.3. shows an accountinf of how Soar would predict the time it takes a person to type dlt . First is the processing that acquires the word and obtains its internal symbol: perceive the sensory stimulus (in the experimental situation, the word was presented on a computer display); encode it (automatically) to obtain its internal symbol; attend to the new input; and comprehend it to be the task word. Second is the cognitive processing that develops the intended answer: getting each syllable; extracting each letter; determining if it is consonant; and, if so, creating the command to the motor system that constitutes the internal intention. Third is the motor processing: decode the command, and move the finger to hit the key (successively d, l , and t ). The entire response is predicted to take about 2.1 sec. (2,140 ms), whereas it actually took 2.4 sec.

Soar is operating here as a detailed chronometric model of what the human does in responding immediately in a speeded situation. This does not fit the usufl view of an AI-like system, which is usually focused on higher-level activities. But a theory of cognition must cover the full tem-

poral range of human activity. In particular, if the theory of the architecture is right, then it must apply at this level of immediate behavior.

These operators and productions are occurring within the architectural frame indicated in figure 9.1. But Soar is not an origional theory here. Lots of psychological research has been done on such immediate-response tasks, both theoretical and experimental. It has been a hallmark of modern cognitive psychology. In this case, the experimental work goes back many years (Fitts and Seeger 1953), and there is an extant theory, developed primarily by Bonnie John (1987), which makes predictions of stimulus-response compatibility. What is being demonstrated is that Soar incorporates the essential characteristics of this theory to produce roughly the same results (the numbers subscripted with bj are the predictions from John's theory).

Acquiring a task

Figure 9.4 shows a sequence of situations. At the top is a variant of a well-known experiment in psycholinguistics from the early 1970s (Clark and Chase 1972). In the top panel, a person faces a display, a warning light turns on, then a sentence appears in the lefthand panel and a picture of a vertical pair of symbols in the right-hand panel. The person is to read the sentence, then examine the picture and say whether the sentence is true or not. This is another immediate-response chronometric experiment, not too different in some ways from the stimulus-response compatibility experiment above. In this case, one can reliably predict how long it takes to do this task, depending on whether the sentence is in affirmative or negative mode, uses above or below , and is actually true or false. This experiment, along with many others, has shed light on how humans comprehend language (Clark and Clark 1977).

Our interest in this example does not rest with the experiment itself but with the next panel down in the figure. This is a set of trial-specific instructions for doing the task. A cognitive theory should not only predict the performance in the experiment but also how the person reads the instructions and becomes organized to do the task. The second panel gives the procedure for doing the task. Actually, there were two variants of the experiment, the one show, and one where 4 reads "Examine the picture" and 5 reads "Then read the sentence." These are not the only instructions needed for doing the task. The bottom who panels indicate increasingly wider contexts within which a person does this task. These panels, written in simple language, are an overly homogeneous and systematic way of indicating these layers of context. In an actual experiment, the person would gather part of this information by observation, part by the gestures and behavior of the experimenter, and part by interaction directly with the experimental apparatus.

The experiment occurs 1. Light turns on. 2. Display shows.

3. Subject roads, exzmines, and press a button. Prior trial-specific instructions 4. "Read the sentnece." 5. "Then examine the picture." 6. {Press the T-button if the sentence is true of the picture." 7. "Push the F-button if the sentnece is false of the picture." 8. "Then the task is done." Prior general instructions 9. "at some moment the light will come on." 10. "After the light comes on, a display will occur." 11. 'The left side of the display shows a sentence." 12. "The right side of the display shows a picture." Introduction 13. "Hello." 14. "This morning we will run an experiment." 15. "Here is the experimental apparatus." 16. . . .

Fig. 9.4.
Acquiring a task

Soar does both the top two panels (but not the buttom two). Focusing on the second panel, as the interesting one for our purposes, Soar takes in each simple sentence and comprehends it. This comprehension results in a data structure in the working memory. Soar then remembers these specifications for how to behave by chunking them away, that is, by performing a task whose objective is to be able to recall this information, in the context of being asked to perform the actual task. On recalling the instructions at performance time, Soar performs the task initially by following the recalled instructions interpretively, essentially by following them as rules. Doing this leads to building additional chunks (since Soar builds chunks to capture all its experiences). On subsequent occasions, these chunks fire and perform the taks without reference to the explicitly expressed rule. Soar has now internalized this task and performs it directly thereafter.

The point is that Soar combines performance and task acquisition in a single theory, as required of a unified theory of cognition. It shows one advantage of having unified theories. The theory of the performance

task is not simply stipulated by the theorist (as Clark and Chase had to do) but flows, in part, from the theory of how the task instructions organize the person to do that performance.

Problem Solving

Let us move up the time scale. Figure 9.5 shows a little arithmetical puzzle called cryptarithmetic. The words DONALD, GERALD, and ROBERT represent three 6-digit numbers. Each letter is to be replaced by a distinct digit (e.g., D and T must each be a digit, say D = 5 and T = 0, but they cannot be the same digit). This replacement must lead to a correct sum, that is, DONALD + GERALD = ROBERT. The figure shows the behavior of a subject solving the puzzle (Newell and Simon 1972). Humans canb e given cryptarithmetic tasks and protocols obtained from transcripts of their verbalizations while they work. The subject proceeds by searching in a problem space; the figure shows the search explicitly, starting in the initial state (the upper left dot). Each short horizontal segment is an operator application, yielding a new state. When the search line ends at the right of a horizontal line, the subject has stopped searching deeper and returns to some prior state already generated (as inicated by the vertical line, so that all vertically connected dots represent the same state on successive returns). The subject often reapplies an earlier operator, as indicated by the double lines, so the same path is retrod repeatedly.

It takes the subject about 2,000 sec. (30 min.) to traverse the 238 states of this search, averaging some 7 sec. per state. Although a puzzle, it is still genuinely free cognitive behavior, constrained only by the demands of the task. This particular data is from 1960, being part of the analysis of problem solving by Herb Simon and me (Newell and Simon 1972). A unified theory of cognition should explain such cognitive behavior, and Soar has been organized to do so, providing detailed simulations of two stretches, lines 1–4 and 8–12. Figure 9.6 shows the more complex behavior fragment (lines 8–12), where the subject has trouble with column 5 of the sum (E + O = O) and thus goes over the material several times, a behavior pattern called progressive deepening . These two stretches are far from the whole protocol, but they still amount to some 200 sec. worth.

The reason for reaching back to old data is the same as with the stimulus-response compatibility and the sentence-comprehension cases. Initially, the most important element in a proposed unified theory of cognition is coverage—taht it can explain what existing theories can do. One attempts to go further, of course. In the sentence case, it is getting the theory to cover the acquisition of the task by instruction. In the cryptarithmetic case, it is attaining completness and detail.

Development

Finally, consider an attempt to understand how the development of cognitive functions might occur. This territory has been

Fig. 9.5.
Behaviro of a person on the cryptarithmetic task

mapped out by Piaget, who gave us an elaborate, but imperfect and incomplete, theoretical story of stages of development, with general processes of assimilation and accommodation, oscillating through repeated equilibrations. Piaget also mapped the territory by means of a large and varied collection of tasks that seem to capture the varying capabilities of

Fig. 9.6.
Soar simulation of the cryptrithmetic task

children as they grow up. Some are widely known, such as the conservation tasks, but there are many others as well.

This exploration with Soar uses the Piagetian task of predicting whether a simple balance beam (like a seesaw with weights at various distances on each side) will balance, titl right, or tilt left with various placements of weights. As they grow up, children show striking differences in their ability to predict, only taking total weight into account (around 5 years), to considering both weights and distance, providing they are separable, to (sometimes) effectively computing the torque (by late adolescence). Developmental psychologists have good information-processing models of each of these stages (Siegler 1976), models that are consonant with cognitive architectures such as Soar. What is still missing—here and throughout developmental psychology—is what the transition mechanisms could be (Sternberg 1984). That, of course, is the crux of the developmental process. It will finally settle, for instance, whether there really are stages or whether cognitive growth is effectively continuous.

Soar provides a possible transition mechanism. It learns to move

through the first two transitions: from level 1 (just weights) to level 2 (weights and distance if the weights are the same) to level 3 (weights, and distance if they do not conflict). It does not learn the final transition to level 4 (computing torques).^[3] Soar predicts how the vbeam will tilt by encoding the balance beam into a description, then using that description to compare the two sides, and finally linking these comparisons to the three possible movements (balance, tilt-left, tilt-right). Soar has to learn both new encodings and new comparings to accomplish the transitions, and it does both through chunking. Figure 9.7 provides a high-level view of the transition from level 1 to level 2. It shows the diffrent problem spaces involved and only indicates schematically the behavior within problem spaces. My purpose, however, is not to show these learnings in detail. In fact, both types of learning are substantially less rich than needed to account for the sorts of explorations and tribulations that children go through.

The above provides the context for noting a critical aspect of this effort to explore development with Soar. Soar must learn new knowledge and skill in the face of existing learned knowledge and skill, which is now wrong. In this developmental sequence, the child has stable ways of predicting the balance beam; they are just wrong. Development implies replacing these wrong ways with correct ways (and doing so repeatedly). That seems obvious enough, except that Soar does not forget its old ways. Chunking is a process that adds recognitional capability, not one that deletes or modifies existing capability. Furthermore, the essence of the decision cycle is to remain open to whatever memory cn provide. Soar, as a theory of human cognition, predicts that humans face this problem, too, and there is good reason and some evidence on this score. Humans do not simply forget and destroy their past, even when proved wrong.

The solution within Soar is to create cascades of problem spaces. If an existing problem space becomes contaminated with bad learning, a new clean space is created to be used in its stead. That is, whenever the old space is to be used, the new one is chosen instead. Of course, when first created, this new space is empty. Any attempt to use it leads to impasses. These impasses are resolved by going back into the old space, which is still around, since nothing ever gets destroyed. This old space contains the knowledge necessary to resolve the impasse. Of course, it also has in it the bad learning. But this aspect can be rejected, even though it cannot be made to go away. The knowledge for this must come from a higher context, which ultimately derives from experimental feedback. Once an impasse has been resolved by appropriate problem solving in the old space, chunks are automatically formed (as always). These chunks transfer this knowledge into the new space. Thus, on subsequent occurrences

Fig. 9.7.
Problem spaces used in learning about the balance beam

of using the new space, it will not have to return to the old space. It may do so for some other aspect, but then that too is transferred into the new space. Gradually, with continued experience, the new space is built up and the old space entered less and less often. But it always remains, because Soar never knows all the information that was encoded in the old space, nor could it evaluate its quality in the abstract. Only in the context of an appropriate task does such knowledge emerge.