Cognitive Elements in Problem Solving
This chapter will bring together and examine work done in the "natural tendencies" areas of decision making research. These areas include cognitive features, and cognitive biases in decision making. These issues will be seen as the basis of our abilities to apply values and to make judgments when making decisions. Some of the classical ideas by scientists in both AI and the DSs are represented here. These scientists include among others, Newell & Simon, Tversky & Kahneman, and G. A. Miller. The chapter will be capped by a look at topical areas in AI and DS which have been developed for supporting weaknesses in the natural tendency area. Cognitive Features The study of man's mind, mental and emotional processes has resided for many years within the discipline of psychology. In more recent years we have seen a break off of this study into an area now known as cognitive science. The concern of cognitive scientists lie in several areas including; Perception Thinking Memory Language & Communication. The focus of the first part of this chapter will be to examine each of these areas.
Why should this study be necessary? The answer is clear. Man has limitations and biases in all these areas which affect his ability to make "good" decisions. Cognitive scientists see as their goal not only the study of mental processes, but also the symbolic modeling of these processes. This points to an eventual goal of overcoming problems in this area. Johnson-Laird[CL3], a cognitive scientist, has stated: "I shall treat the study of the mind as a scientific end in itself, but the resulting knowledge has already contributed to the practice of teaching and learning, to the diagnosis and treatment of mental disorders, and to the design of a humane technology that augments human intelligence. "There are ample examples where cognitive limitations and biases have resulted in disastrous decision making[CB6]. Additionally, considerable research has been devoted to pointing out the repeatability of these deficiencies. See Edwards and von Winterfeldt for a good review of research. Some of these limitations and biases will be reviewed here from the standpoint of the 4 categories listed above.
Perception has a key role in the cognitive domain. Perception is defined by Schermerhorn et al [CL6] as ".. a screen or filter through which information must pass to achieve impact on human thought processes and behaviors." The figure below provides a more visual sense of perception.
Figure 3-1. Perception
Perception may be the single largest area where cognitive differences exist between humans. We each have very different natural abilities and learned experiences (developed filters) which cause our perceptions of things and events to be very different from one another.
As a simple example of differences in natural abilities, what is it that makes a professional baseball player a consistently good batter? What is the basis of their decision about when to swing? Doubtless it is the combination of many things including attitude, practice, coaching, and a well formed understanding of the game. But it is also a fact that a batter in professional baseball has less .5 seconds to respond to a ball being thrown toward them at speeds nearing 100 MPH. Discerning what is happening with the ball is not a task which the average eye is capable of performing, irrespective of training. Outstanding batters such as Wade Boggs and Pete Rose have been shown to have clearly superior eye strength and tone. We all exhibit varying degrees of strength and weakness in this and the other sensory areas hearing, taste, smell, and touch. These natural abilities all have a bearing upon our lifelong pursuit of problem solving and decision making. Gregory [DA1] provided a case where a man, who had been blind since the age of 10, suddenly had his eyesight restored at age 52. When shown different objects, he had no idea or recollection of what they were. But upon closing his eyes and feeling even a small part of those same objects he could identify them immediately. His sense of touch had been magnified to compensate for deficiencies elsewhere. While natural ability has a definite impact upon our perception abilities, the primary factor in perception are differences in learning.
Learned Experiences & Perception
Klein [CL4] and others have done some interesting work with firefighters. Klein observed that upon arriving at the scene of a blaze, an experienced firefighter could quickly appraise the situation and immediately begin acting to mitigate the potential damage. In addition, experienced firefighters seemed to work in complete concert with each other, without having to pass a communication amongst themselves. On the other hand, rookies, even after thorough training, were unable to perceive where to begin. In any profession the difference between a mediocre performance and a valued one is the ability to quickly size up a situation and respond accordingly. Why is it that grand master chess players can play an entire room full of novices and yet still win consistently? It is because through learned experience they can recognize distinct patterns in the positions of the chess pieces, and can react with habitual precision to attack an opponent or to counter an attack.
Attention in Perception
Some researchers include attention as a separate part of perception. Attention can be seen to be those attributes of perception concerned with focus, depth and duration. For example why is it that some events in life catch our attention and others do not? And why is is that some events catch our attention to the detriment of others including our overall well being? Take the situation of a wasp landing on your nose while you are driving down a crowded highway? The natural reaction is to swat it. But losing control of your vehicle would pose a much greater threat to your well being than a wasp sting. This is a case of misplaced attention. On the other hand, have you ever mentally drifted off while on a long interstate drive, only to "suddenly" awake to the immediacy of the road and your driving? Where did the last 5 miles go? This is a case of unfocused attention or perhaps focused subconscious perception.
Some Examples in Perception and Decision Making
There have been a number of interesting experiments performed in perception as it relates to decision making. Schermerhorn, Hunt & Osborn have detailed a few in Chapter 14 of their book Managing Organizational Behavior[CL6]. While not an experiment, the drawing below demonstrates how an identical situation may be viewed from very different angles. Do you see the old lady? the young lady? both?
Figure 3-2. Portrait of a Lady
While Portrait of a Lady provides a superficial look at differences in perception, a study of management/employee attitudes toward identical behavior points out how day to day operations can also be affected. The study demonstrated how polar perceptions can be about what is occurring in the workplace. This is a common situation. Management points at problems created by employees while employees point at poor management practices.
Figure 3-3. Managerial v. Subordinate Perceptions
Many more cases of how differences in perception affect DM abound. This is a classic problem which dates far back in time. Recall the fable of the 6 blind men and the elephant, each thinking they had identified the object as a rock, a snake, a tree trunk, etc. Perception plays a large part in the elements of cognitive style and DM. Humans have been likened to limited capacity information processing machines that have processing capabilities on several levels. Once the perceptive filtering has had a chance to "weed out" the large quantity and quality of information received by the brain, the remaining data must be processed. This processing, in the world of cognitive science, is termed thinking or reasoning.
What is reasoning? Philosophers and scientists have spent considerable time looking at what occurs after data enters the brain and before the corporeal body is moved to respond. Consciously and subconsciously humans combine perceived data with memorized data to formulate conclusions. Several distinct formal reasoning approaches have been formulated and studied. These include; inductive, deductive, analogical, procedural, and meta-level.
Inductive reasoning is the approach utilized when no past experience exists in the area of the collected data. A single example is collected and generalized. As new examples are discovered, and the rewards system for reacting to these examples provides positive or negative reinforcement, our reasoning process changes. Eventually when sufficient material is experienced, the deductive process becomes dominant. The deductive method begins with general principles, attempting to bind immediate facts to general situations, and to act or not act according to the reward paths carved from past experience. The inductive process is only again invoked when the deductive process cannot properly accommodate new data. Analogical reasoning runs parallel to deductive and inductive reasoning, overriding those processes when appropriate. It is the process of looking for a similar or identical case. Analogical reasoning can be seen to be very effective reasoning. It masks out unimportant data, locating "close" cases quickly. The procedural approach processes data sequentially using an algorithmic approach. For most tasks it is a cumbersome method. It is computationally inefficient but may be seen as effective in situations where complexity and confusion are particularly evident. Finally meta-level reasoning is that cognitive mechanism which controls our reasoning about reasoning. For example we may begin a reasoning process by searching for similar cases. But when the search space becomes too large or time consuming to satisfy our practical constraints, we may switch to a deductive mechanism. Some typical characteristics of meta-level reasoning include ordering, restricting, pruning and canonicalization. Reasoning and the representation of the reasoning process is a primary focus within the field of artificial intelligence and will be examined again later in this paper.
Another large part of cognitive science is memory. Cognitive dissonance occurs commonly in human memory. This is one reason why a large part of intelligence testing is the measurement of memory capabilities. Psychologists measure both short term and long term memory to determine the limits of an individual's abilities. Short Term Memory G.A. Miller, in a classic article[CL5] reviewed experiments involving auditory, taste, visual and mental capacities.
"There is a clear and definite limit to the accuracy with which we can identify absolutely the magnitude of a uni-dimensional stimulus variable. I would propose to call this limit the span of absolute judgment, and I maintain that for uni-dimensional judgments this span is usually somewhere in the neighborhood of seven."
Miller's thesis is easily duplicable by asking a subject to memorize a series of seven single digit numbers spoken aloud to them, and then to repeat the numbers a minute later. This recall process tests short term memory. Facts and events which are often repeated, or emphasized through an especially strong sensory experience gradually move from short term into long term memory. Memory therefore is seen to exist on a continuum from immediate to lifelong.
Long Term Memory
An interesting cognitive study on problems inherent in long term memory and decision making was provided by Jenkins[CL2]. Jenkins analyzed the infamous Watergate tapes of the Nixon administration. What prompted Jenkin's research was the seemingly contradictory Capitol Hill testimony provided by administration members several months after the supposed "misdeeds". What struck him odd was the interplay between 1) an individual's testimony about a conversation, 2) the testimony of his colleagues about the same conversation, and 3) the actual conversation as provided by the tapes. While there were frequently large differences in interpretations of the conversations, none of the participants could be said to be lying about their part. The large differences were simply a matter of memory and viewpoint. As Jenkins states: "Recent research has been showing that memory is much more than a passive recorder of events. Instead, our minds seem to have a need for coherence. This is so much the case that when events are piecemeal and lack sufficient cohesion, our minds strive to fill in the missing pieces to give meaning to the stream of information. It is finally this meaningful whole that is stored most strongly in memory, not the unconnected pieces." This research has significant implications for problem solving and for decision making. What we frequently discern as "fact" may differ widely based upon our background, and frame of reference. Beyond this we must also be cognizant that the ability to memorize "facts" differs from retrieving them in an unbiased manner, accurately, or within a reasonable period of time. There are many bias problems in this arena. They will be addressed in the later section on cognitive biases. While perception, thinking and memory all evolved and work as inputs and processes within our brains, our final cognitive factor, language, works external to our introspective cognitive domain and therefore is the one area most likely to be misunderstood and misinterpreted.
Language & Communication
The nature of human to human communication is a complex one having evolved over many millenniums. Spoken communication is estimated to have originated over 50,000 years ago with written communication begun an estimated 5,000 years ago. Research in this cognitive domain has proceeded now only for a few decades largely under the heading of linguistics. Researchers have hypothesized the structure and meaning of language by creating a tiered approach for its study. The layers in this tier include; phonetic representation, morphological analysis, lexical complications, syntax, semantics, pragmatic indicators, and even "real world" modeling. There can be no doubt that many of the most horrific actions in the history of man have been started with the misinterpretation of language. So the following paragraphs look at the above seven levels with an eye toward better understanding how this area may be applied toward creating better models for decision making. Phonemes are those smallest building blocks of language upon which words and sentences are built. They are the foundations of the literal spoken word. Using these small building blocks technologists have been able to build successful speech recognition devices. But being able to recognize speech is a miniscule task compared to being able to "understand" it.
Morphology, information about words and their root forms is a first step. For example the sentence, "We're sailing to Rhode Island" must be taken apart to understand that "we're" is actually "we are" and that "sail" is the root form of "sailing". The second hurdle is that of lexical complications, or the basic "dictionary" meaning of words. For example, what does "Stay away from the bank" mean? There are many types of banks including river banks and banks where money is stored. Problems with syntax, which deals largely with grammatical rules, is the next hurdle in understanding. It deals with such subjects as correct tense, and proper sentence structure for subject verb agreement, etc. The next hurdle in understanding is semantics, or the underlying meaning in a communicator's speech. For example take the sentence "The chickens are ready to eat." Who is ready to eat? The family or the chickens? Even beyond semantic questions are pragmatic indicators. For example the sentence "She saw the bicycle in the window and she wanted to buy it." Did she want to buy the window or the bicycle? We understand she wants to buy the bicycle. But we only understand that because we have "common sense" about seeing bicycles in windows. We cannot hope to understand language fully until all factors behind that language come into play. And last but probably most important are real world or reasoning models. This capability is the finesse to look far beyond what may be implicit in words to match what is being said to some real world experience. This may include the ability to verify the reasoning process in a statement, or to verify the existence of an imperative or the validity of a question.
We have now examined the 4 basic elements of cognition perception, thinking, memory, and language. Humans possess vast skills in each of these areas. But humans are not "all seeing, all knowing or all wise" creatures. Many problems exist in the human use of these cognitive elements. The following section examines some of these problems.
Problems/Biases in Cognition
The scientific study of human weaknesses in decision making has proceeded at a linear pace thru the last 3 to 4 decades. Some of the major researchers have become very well known for their work. These include A. Tversky, D. Kahnemann, and D. von Winterfeld. This research has been framed in many forms including the study of uncertainty, the assignment of preference values, the interpretation of probability, and the study of "rational" behavior. The following section will look at these factors and others from the viewpoint of the four cognitive areas already covered.
The following list outlines some of the more apparent problems in humans dealing effectively with what is being perceived:
Insufficient Information Collection
Important Data being filtered out
Important Data being channeled erroneously
Poor Source & Quality of Data Data/
Information outside previous scope
Using an Absolute measure v. Relative
Most of these issues are self explanatory. Some have become more prevalent in the information age. Having much greater access to data has made us more aware of them - even to the point where we become aware that the data is lacking in quantity and quality.
The various reasoning approaches were addressed earlier in this chapter. Most of the problem issues below come to light when reasoning is either being applied improperly or being extended beyond its "reasonable" limits, and include;
Escalation Bias due to the source
Misconception of chance
Escalation is the tendency to commit additional resources to failing efforts. Illusory correlation is the assignment of cause and effect relationships where it is unsubstantiated. For example during a short period in the early 1980's it was shown that sun spot activity correlated well with the S&P stock index. Does this indicate that sunspot activity causes the stock market to rise? Misconception of chance is another area where poor thought processes may lead to incorrect conclusions. In the toss of a coin, the probability of it landing heads is always .5, even though it may have landed tails for the last 100 tosses.
In Use of Memory
Fallacies in the use of memory have been studied extensively. Some of these issues follow:
Using an Absolute measure v. Relative
Looking solely for Confirmation
Incorrect Anchoring & Adjustment
Insensitivity to prior probabilities
Insensitivity to sample size
Bias due to retrievability
Bias due to imaginability
As was pointed out in the Watergate example earlier, our brains tend to play tricks with our memories. After we make a decision we tend to only look for confirming evidence that our decision was the "best choice"." Or when asked for evidence on a matter we access only that data which is closest to the surface. Or we only imagine the best or the worst in a situation. And we often ignore small amounts of factual data while placing large amounts of credence in hearsay evidence or irrelevant past experience.
Relating to Language & Communication
Areas of problem origination in language focus on those levels covered earlier and may include;
Anaphora and definite noun phrases
Extra grammatical utterances
Ambiguity exists on many of the levels including lexical, syntactic, and semantic. These range from a simple dual interpretation of an acronym to a variety of word interpretation, e.g. the word "BAD" which in some lingo means "GOOD". An interesting study on background framing was performed by Kent[CB3]. He did a study on words of estimated probability as they relate to uncertainty. The diagram below depicts the results of his study.
Figure 3-4. Words of Estimated Probability
Anaphora and definite noun phrases refer to concepts or phrases referenced earlier in a dialogue. A particular earlier concept or phrase referenced may not always be evident to all parties, and therefore subject to varying interpretation. Ellipsis, extra-grammatical utterances, and meta-linguistic utterances all revolve around informal speech short phrases, interjections, false starts and "Oh, I meant to say .." sorts of problems.
It should not be construed from pointing out the preceding cognitive problem areas that man is doomed for "decision making extinction". Weaknesses in some areas are often supplanted by overwhelming strengths in others. An individual with a weak memory may substitute a strong language ability. A person with a weakness in perception might compensate with exceptional abilities to reason. Nonetheless most would agree that "good" decisions require strength in all areas. Therefore decision scientists and AI researchers are looking for techniques and "machines" to support our processes in those areas where weaknesses do exist. The following table outlines some areas within the AI, DS and Computer Science disciplines where support is evident.
|Thinking / Reasoning|
║ Cognitive ║ Distinct Support Mechanisms ║ ║ Area AI ║ DS ║ Computing ║ ╠ ║Perception ║ Neural Networks; ║ Kalman Filtering; ║ Constant Vigilance║ ║ ║ Rules; Induction; ║ ║ ║ ║ ║ Certainty Factors ║ ║ ║ ║Memory ║ Large KBS; Frames ║ Modeling ║ Large Data Storage║ ║║Thinking ║ Reasoning Methods; ║ MCDM Methods; ║ ║ ║ ║ Truth Maintenance ║ Statistics; ║ Processing Power ║ ║ ║ Systems ║ Resource Allocation║ ║ ║Language ║ Natural Language ║ Frameworks ║ Programming ║
Table 3-1. De-biasing Methods
Problem solving, decision making, and the cognitive limitations and biases areas have now been reviewed. It is now time to move on to address those areas within Artificial Intelligence and within the Decision Sciences designed to address these areas.