by James Reason
4.1. Similarity-matching
As indicated earlier, this process entails the continuous matching of the recent output of FWM (termed the ‘calling conditions’) to attributes of knowledge units. To avoid exhaustive searches of the entire knowledge base, the first ‘pass’ is directed at those stored units sharing similar contextual elements to the recent contents of FWM (i.e., those currently in the KB’s buffer store). Such elements may be said to define the contextual frame of the search. If this first pass fails to produce a satisfactory match, then other frames must be found through FWM activity.
Note also that retrieval is initiated automatically by the products of FWM activity. No special fiat or directive is required. But the products of this automatic similarity-matching need not necessarily arrive in FWM. They may be excluded by the selection criteria currently operative within PWM, or they may simply be pre-empted by higher priority inputs. The greater the activational charge of the knowledge units emitting the search products, the greater is their chance of reaching FWM.
4.2. Frequency-gambling
In many situations, the calling conditions emerging from FWM are insufficient to provide a unique match for a single knowledge unit, because either the calling conditions or the stored elements are incomplete. These two possibilities are functionally equivalent. When searches are underspecified, a number of partially matched ‘candidates’ may arrive in PWM. Where these contenders are equally matched to the calling conditions, the conflict is resolved in favour of the most frequently-encountered item by the activation principle. An oft-triggered knowledge unit will have a higher ‘background’ activation level than one that is less frequently employed. This is generally an adaptive heuristic because the most contextually typical search products will be dealt with first by FWM. In other words, the machine responds to the statistical properties of the world it inhabits.
4.3. Directed search
FWM has no direct access to the stored knowledge units, only to their products. Its sole means of directing knowledge retrieval is through the manipulation of calling conditions. The actual search itself is carried out automatically by the similarity-matching and frequency-gambling heuristics. All that FWM can do, therefore, is to deliver the initial calling conditions, assess whether the search-product is appropriate and, if not, to reinstate the search with revised retrieval cues. FWM thus has the power to reject the high-frequency candidates thrown up by underspecified matches, but only when sufficient processing resources are available.
5. Intentional characteristics
As we have seen, actions can be set in train and knowledge products propelled into FWM as an automatic consequence of (a) prior processing in FWM and (b) knowledge unit activation. But these properties alone are not sufficient to guarantee the successful execution of goal-directed behaviour. What gives the machine its intentional character? How does it initiate deliberate actions or knowledge searches?
As with many other questions, William James (1890, p. 561) provided an answer: “The essential achievement of the will... is to attend to a difficult object and hold it fast before the mind. The so doing is the fiat; and it is a mere physiological incident that when the object is thus attended to, immediate motor consequences should ensue.”
This statement maps readily onto the properties of our machine. The ‘holding-fast-before-the-mind’ translates into a sustained run of same-element FWM slices (see Figure 5.2.). Once in the KB buffer, the consistency of these slices will generate a high level of focused activation within a restricted set of knowledge units. This will automatically and advantageously release their products, either to the output function or to WM. But such an ‘act of will’ places heavy demands upon the limited attentional resources available to working memory. The perseveration of specific elements has to be maintained in the face of continual pressure from other claimants to FWM attention. And such an ‘effort’ can only be sustained for short periods, particularly when the object being attended to is less than compelling. As James (1908, p. 101) put it: “When we are studying an uninteresting subject, if our mind tends to wander, we have to bring back our attention every now and then by using distinct pulses of effort, which revivify the topic for a moment, the mind then running on for a certain number of seconds or minutes with spontaneous interest, until again some intercurrent idea captures it and takes it off.” like the human mind, the ‘fallible machine’ is naturally prone to “intercurrent ideas.”
As the machine’s recurrent actions become compiled into preprogrammed instructional sequences, FWM need only focus upon (i.e., sustain an image of) the desired consequences, not upon the detailed movements. To overcome the difficulty of having the effects of an action precede the action itself, James’s ideomotor theory maintained that the imagined consequences of voluntary movements are derived originally from the experience of reflex behavioural units: “When a particular movement, having once occurred in a random, reflex or involuntary way, has left an image of itself in the memory, then the movement can be desired again, proposed as an end, and deliberately willed” (1890, p. 487). Similarly, our machine could begin its working life with a limited stock of ‘hard-wired’ action, ideational and perceptual programs. Acquired knowledge units would develop as increasingly elaborated variants of this basic ‘starter-kit’.
6. Concurrent processing
The principal achievement of human cognition is its extraordinary ability to internalise the recurrences of the world in long-term memory (as schemata, scripts, frames, rules and so on), and then to bring the products of these stored knowledge structures into play whenever the current situational calling conditions demand them. What is even more remarkable is that these retrieval WORKING MEMORY: Contents of focal WM are “sliced” and each slice (corresponding to a processing cycle of a few milliseconds) is dropped Into the KB buffer store activities are performed in a largely automatic fashion, without recourse to the computationally powerful yet highly restricted operations of the conscious workspace (or working memory).
Figure 5.2. The role of intention in the fallible machine. Intentional activity involves the allocation of limited attentional resources to sustaining a run of same-element slices within FWM. The consistency of these slices creates a high level of focused activation within a restricted set of knowledge structures. This, in turn, helps to preserve the coherence of FWM activity.
This is not to say that conscious operations play no part in knowledge retrieval, simply that long-term memory is capable of emitting its products (actions, thoughts, images, words, etc.) without the necessity of such higher-level direction. Indeed, it has been argued here that the conscious workspace has no more privileged access to its associated knowledge base than do inputs from the world at large. The difference lies not in the degree of access, but in its specificity. Whereas the conscious workspace puts no constraints upon the type of information it will accept, so long as it lies within the scope of the sensory system to detect it, each knowledge structure within long-term memory is tuned to a highly specific set of triggering conditions and is largely oblivious to all that falls outside this exceedingly narrow range. Thus, while the conscious workspace keeps ‘open house’ to all kinds of information, either externally or internally generated, the knowledge base constitutes a vast community of specialists, each one scanning the world for only those inputs that match its own very parochial concerns.
A key feature of this machine, then, is that information is processed simultaneously by both working memory and the knowledge base. The concurrent operations of these two structures are shown diagrammatically in Figure 5.3.
Theoretically, this interchange between the automatic search processes of the KB and the inference ‘work’ carried out by FWM could continue until an appropriate solution has been found. In practice, however, the machine will probably ‘home in’ upon a hypothesis quite early on. Thereafter, that hypothesis is likely to be sustained because the similarity-matching is tailored to finding confirmatory rather than disconfirmator
y evidence. Such a search also minimises cognitive strain.
When faced with a simple assembly task, for instance, many people adopt the strategy shown in Figure 5.4. That is, instead of reading the instructions carefully and then following them step-by-step, they frequently just ‘fiddle’ with the problem configuration until a recognisable pattern (idea) comes along, and then act upon it. This creates a new configuration... and so on. In other words, they act (a) to minimise cognitive strain and (b) to maximise the chances of automatic pattern recognition. Likewise, children (or adults), when set an apparently difficult problem will search their questioner’s face, the situation or other people’s expressions for a clue to act upon, even though they have been provided with sufficient information to derive the solution deductively.
Figure 5.3. Concurrent processing by the knowledge base and working memory during problem solution. Step 1. The first solution is delivered to WM as the result of automatic similarity-matching and frequency-gambling acting upon the initial retrieval cues. Step 2. The solution is evaluated by WM and found inadequate. Analytical processes generate revised retrieval cues. These are processed automatically within the KB, and a second solution is delivered to WM. Step 3. Again the solution is judged as inadequate, and further cues are generated by WM processing. Step 4. The third solution is found to be satisfactory and emitted.
Figure 5.4. An everyday problem-solving strategy. Instead of doing inference work, default solutions are continuously emitted until a familiar pattern comes along. The problem is worked out in action rather than in thought.
7. Taking stock
Any attempt to model the fundamentals of human cognition must address two basic issues: (a) the properties of the knowledge base and its modes of representation, and (b) a set of rules or heuristics for selecting which stored knowledge structure will be activated in any given situation. This response-selection element not only provides the model with its human information-handling characteristics, but also creates and shapes the recognisably human error forms.
In an ideal world, each particular problem configuration would elicit one appropriate stored solution. But the reality is usually far from ideal and contains two major sources of ‘fuzziness’: (a) The calling conditions associated with any particular problem configuration can match several stored structures or none at all; or they can be degraded, not attended to, or absent and (b) knowledge structures can be incomplete (not all facts known), wrong or missing altogether. As far as the cognitive system is concerned, these two kinds of ‘fuzziness’ are functionally equivalent. Both represent sources of underspecification (see Chapter 4). These varieties of underspecification create the conditions under which similarity and frequency biases are most likely to show themselves.
Inevitably, this design for a fallible machine dodges many crucial issues and skates over others. In particular, it says very little about inference, and nothing at all about the way FWM decides that it has found an appropriate problem solution. What it tries to do, however, is to convey a picture of an information-handling ‘machine’ that, though capable of internalising the complexity of the world around it, is in essence driven by a small number of simple computational principles. In the next part of this chapter, we will examine one attempt to implement these notions within a computer model designed to simulate the retrieval of incomplete knowledge.
8. Modelling the retrieval of Incomplete knowledge
8.1. Category generation
This section describes the basic features of a computer model (implemented in Prolog) designed to emulate the way in which people with a relatively sparse knowledge of United States presidents respond to the request to generate exemplars of that category. All the computer models described in this chapter were written by Philip Marsden of the University of Manchester.
8.1.1. What the model ‘knows’
The model’s knowledge base comprised 32 frames, corresponding to each of 32 individuals who have occupied the office of president of the United States, from Washington through to Reagan. (Why this should be 32 rather than the (then) complete list of 39 presidents will be explained later.) Each frame contained two pieces of information: (a) the president’s name, and (b) an actual frequency of encounter value (ACTFOE).
The ACTFOE values were obtained from a diary study (Marsden, 1987) in which 10 volunteers kept a daily tally over a period of 13 weeks of the number of times they encountered (in the media, in books, in conversation, etc.) any one of the 39 presidential names. The ACTFOE value for each president constituted the sum of all the encounters over the record-keeping period for all diarists. These ACTFOE values are listed in rank order in Table 5.1.
The following presidents were not logged in the diary study and consequently did not appear in the knowledge base for the category generation simulation: Harding, Madison, Fillmore, Pierce, Andrew Johnson and Zachary Taylor.
8.1.2. How the model works
On each run, the program was designed to simulate the output of a single human subject when the latter was asked to generate as many names of American presidents as possible within a time span of approximately 5 minutes. The structure of the program is most conveniently described in three sections:
Table 5.1. Actual frequency of encounter (ACTFOE) values for U.S. presidents.
Ronald Reagan: 1822
Calvin Coolidge: 11
John F.Kennedy: 176
Thomas Jefferson: 10
Richard Nixon: 145
Benjamin Harrison:
Abraham Lincoln: 93
John Tyler: 6
Franklin D. Roosevelt: 75
John Adams: 6
George Washington: 63
William H. Harrison: 4
Lyndon B. Johnson: 61
John Quincy Adams: 3
Jimmy Carter: 57
James A. Garfield: 3
Woodrow Wilson: 56
Rutherford B. Hayes: 3
Dwight D. Eisenhower: 53
William McKinley: 3
Theodore Roosevelt: 47
James K. Polk: 3
Gerald Ford: 43
Grover Cleveland: 2
Harry S. Truman: 40
James Buchanan: 2
Ulysses S. Grant: 31
William H. Taft: 2
Herbert Hoover: 21
Martin van Buren: 1
Chester A. Arthur: 1
Section 1: This is the control section and contains the executive rule for generating presidential exemplars. Ultimately the rule fails; but, in the process, it outputs the results of the search as a side-product.
Section 2: This section converts the ACTFOE values into calculated frequencies-of-encounter (CALCFOEs). The CALCFOE is a random number between unity and the ACTFOE value for each president. These values differ on each run. However, the larger the original ACTFOE, the greater the chance that the CALCFOE will also be relatively large.
Section 3: This section deals with the knowledge base search procedures, and contains the following components:
(a) search is the mechanism that guides the scanning process. It has three parameters: upper (the upper limit of the search scan—defined as a CALCFOE value), lower (the lower limit) and time (a crude time interval flag). Following each instantiation of scan, search checks to see if the noise threshold (a CALCFOE value below which no presidents are retrieved—analogous to the subject’s knowledge level) has been reached before altering the scan parameters and adjusting the time interval. It then initiates a further search.
(b) set band is a sequence which sets the width of the search process in CALCFOE units. With each successive search, the band becomes narrower, thus retrieving progressively fewer items. This simulates the increasing difficulty experienced by subjects in generating new exemplars as the output progresses.
(c) check band checks the new bandwidth to ensure that its value never falls below unity.
(d) scan is the search band. It checks the presidential knowledge base to establish if the CALCFOE falls within its range. If
it does, then it returns the president’s name, the CALCFOE value and the time interval.
(e) set noise establishes a different noise threshold for each run. This varies randomly between 11 and 49 CALCFOE units and is intended to simulate the fact that within a given population of subjects there will be differences in the level of knowledge (i.e., a value of 11 corresponds to relatively high knowledge, and a value of 49 represents a very low knowledge level).
(f) noise is the noise threshold, a value (in CALCFOE units) below which no items can be retrieved.
The model thus has two sources of quasi-random fluctuation from run to run: the variable CALCFOE value for each president within the knowledge base, and the noise threshold, varying between 11 and 49 CALCFOE units. Since it scans downwards from high CALCFOE values to lower ones in increasingly smaller ‘bites’, it will show a marked but not invariant tendency to output high CALCFOE presidents first. This, in turn, will be positively (but not perfectly correlated) with the ACTFOE values. That is, there is a built-in tendency for items to be generated in an order corresponding roughly to presidential frequency of encounter in the world, as determined by the diary study. This tendency corresponds to well-established findings (Bousfield & Barclay, 1950; Battig & Montague, 1969; Reason, 1984,1986).
8.1.3. The model’s output
