Inferred Functions of Performance and Learning

Siegfried Engelmann, Donald Steely

• Publisher: Psychology Press
• Year: 2003
• eBook ISBN: 9781135624835
• Topic: Psychology
• Subtopic: Cognitive Psychology & Cognition
• Index: Psychology

This ambitious, highly theoretical book provides a capstone for the careers of two very distinguished scholars. It begins with an analysis of what functions and systems must exist for any organism or machine to perform an unlearned act, that is, with an analysis of what must be "wired into" the organism or machine. Once the basics of unlearned responding have been established, the authors then systematically show how learning mechanisms can be layered onto that foundation in ways that account for the performance of new, learned operations that eventually culminate in the acquisition of higher-order operations that involve concepts and language.

This work is of interest to various practitioners engaged in analyzing and creating behavior: the ethnologist, the instructional designer, the learning psychologist, the physiologist-neurobiologist, and particularly the designer of intelligent machines.

Part I

Performance of Nonlearning Systems

Chapter 1

A Framework for the Fundamentals of Performance

The behavior of organisms is purposeful. What they do on any given occasion may be conceived of as a task performed to bring about a functional effect. To perform the tasks as they do, the operating systems of organisms must carry out certain basic functions. A function is that which is essential or common to all possible systems that could perform this task. One operating system may do it one way and a second another way. Yet if there are 100 different system designs that perform the same task, all possess a common set of functions. These functions are identified by analyzing the tasks and specifying the details that are essential for performance, regardless of the specific way in which these functions are carried out.

If a particular task calls for responding to a visual stimulus, all systems that perform that task must be able to meet the functional requirement of being able to receive visual information. The specific design of the receptor system may vary from a simple eye to a compound eye or even to a zoom lens. So long as the system has some basis for receiving visual signals, it meets one of the functional requirements imposed by the task.

If an organism or machine responds in a predictable way to a particular visual signal, the system must have some way to associate or connect the perceived visual stimulus to the motor-response system. Essentially, the system needs some sort of rule that alerts the organism to the fact that the presence of the particular stimulus requires a particular response.

If an organism or machine has a repertoire of different responses for various visual signals, it must have some form of screening function that is able to perceive and differentiate the various stimulus inputs and connect them to different response outputs. This requirement is fundamental if the organism (a) produces different responses to a given stimulus under different conditions, or (b) produces the same response in the presence of two or more different visual stimuli. Some type of decision making is implied to determine which response to produce for any particular stimulus because a particular response is no longer a simple function of the stimulus.

The analysis of functions does not need to consider neurology, anatomy, physiology, biochemistry, genetics, or psychology in identifying these basic functions. Rather, it considers only the logically implied, universal features that would be identified in each of the 100 different systems designed to perform the same task.

In working on this analysis, we repeatedly identified functions that seemed far too sophisticated for the organisms that produced the behavior. For example, the turning point in formulating the present version of the work occurred when we had begun a serious logical assault on what functions are needed for a single-cell organism to identify and approach a food source. The task we addressed was that of approaching a source of olfaction by using only olfactory sensory input. A casual analysis suggests that all the performer had to do was keep going in the direction that made the perception of the olfactory stimulus stronger and stronger until reaching the source.

However, a more careful analysis of the task revealed that achieving this outcome requires an intricate interaction between information and logic. We consider only one aspect of this interaction for the time being, which is that the system must be able to perform some form of temporal analysis. Specifically, the organism must do a comparison of the levels of olfactory receptions at two separate points in time and then determine that at one of the times the stimulus is stronger than it was at the other time. Without this information, the system would not know whether it is should proceed in the direction it had been going or move in another direction. Furthermore, the comparison of the data at Time 1 and Time 2 cannot be performed until the second sample is obtained at Time 2. Therefore, the information from Time 1 must have been stored so that it is available to the organism as some form of representation. Once the organism concludes that reception at Time 2 is greater than that at Time 1, the system must produce behavior that is consistent with the conclusion. For all occasions in which the Time 2 reception is greater than the Time 1 reception, the behavior was successful, so the directive that results is functionally equivalent to the verbal directive, “Keep doing behavior X.”

Theory of Inferred Functions

In some ways, a theory of inferred functions proceeds in a different direction than current theories of learning or intelligence. The main difference is that inferred functions address more details of the content and logic of the system that performs. Questions of the stimulus become questions of the features of the stimulus and the steps that would be required for the organism to be able to recognize the stimulus on various occasions, although the precise pattern of stimulation is never repeated. Questions about the response become questions about the kind of information the system needs on a given occasion to plan a response, direct it, and adjust it on the basis of feedback.

Although the analysis of inferred functions leads to a greater articulation of content and logic, the analysis is an extension of behavioral analysis. Behavioral analysis reveals the functional relationships between stimulus events and responses—how behavior is affected by changes in discriminative stimuli and the consequences of behavior. The present analysis of inferred functions takes the next step in identifying the internal functions implied by the stimulus–response relationship.

One contraintuitive aspect of the analysis is that it focuses on the things that the organism encounters—the shapes, colors, changes, and specific features that distinguish each from other things in the surroundings. The analysis of this aspect of performance is contraintuitive because it involves a sometimes technical consideration of the features or stimulus elements that various things and events possess. The analysis leads to issues about how an organism is able to identify a novel stimulus instance as a positive or negative example of a discriminative stimulus when the organism has never encountered that example before.

There are certain stimulus elements in the natural environment to which an organism may innately respond. However, unless the organism has sufficient knowledge of those things to identify them and has knowledge of those features of a response strategy that change the current setting in specific ways, the organism cannot perform. Just as an organism’s system has been shaped to exploit the surroundings by using sophisticated receptors and performing sophisticated chemical analyses, it must conduct an equally sophisticated analysis of the specific features—properties, behaviors, tendencies—of the things that are important to the organism’s survival.

The bottom line is that all the basic discriminations and strategies the organism has or learns are perfectly logical because things in the surroundings follow rules that are governed by logic. If an object is here, it is not there, and the system that performs must record where it is. If the object is transformed in universal ways when movement occurs (e.g., becoming apparently larger as it is approached), the system that performs must have provisions for accommodating the effect of the transformations and recognizing what remains the same about the object. In the same way, if the object differs from something else in three specific features, the learner must learn the nature of these features.

Furthermore, the learner must have all the logic required to draw conclusions about which features of the object are correlated with reinforcement. The logic is needed to determine what the presence of these features predicts. Without such logic, there can be no performance or learning because there would be no way for the organism to interact with the things in the surroundings. Whatever the organism learns is data based. There are no general, nondistinct learning tendencies. There is only the learning based on facts about the things that are relevant to what is learned.

A Performance Framework

The required components of a performance system may be identified by starting with a simple machine and noting the changes in functions that result when the performance task becomes more demanding and more like that of an organism. Machines can be deceptive because their design permits single parts to perform multiple functions. For the machines that we describe, we separate the functions.

Binary Performance

The simplest machine is designed to produce one response to a stimulus. This machine is an analogue to a spinal reflex. For our machine, a particular visible object—a ball—causes the machine to produce a single response—moving an arm up in a specified arc at a specified rate. This machine requires a minimum of five functions. Some of these functions are obvious and some are not:

1. Reception
2. Screening
3. Planning
4. Directive
5. Response

Reception Function. The reception function is simply the capacity of the machine to receive the type of sensation that contains the positive examples (as well as the negative examples). In this case, the sensation is visual, and the positive example is the ball.

Screening Function. The screening function occurs next. It is essential because sensory receptions contain both positive and negative examples (balls and not balls). The system must have some qualitative criterion for identifying or screening the positive examples and only the positive examples. The criterion rule may take many different forms, but all have a single content function—to identify objects that share specific features. For the screening function, one particular pattern of features is identified as positive; the others (by default) are negatives. Without this screening function, the system would not know when to plan the response. The screening must function as an independent variable and the response planning as a dependent variable.

Planning Function. The planning function is activated when the presence of a positive example has been confirmed. Note that the planning function is not directly linked to either the reception or response. It is the link between the screening function and directive function. Receptions of negative examples result in a default plan to do nothing. Receptions of positive examples result in establishing a specific, planned response.

The planning function does not produce or even direct the response. It simply identifies what the response directive will be. When the plan is activated, it specifies exactly what the machine will do next.

Directive Function. The directive function directs the motor-response system to activate the response. The response has been planned. The directive function tells the system, in effect, “Do that response.” The directive function is not the response, but the process that orders the execution of the response that has been planned.

Response Function. Unlike the other functions, the response function is observed as an action. It is not information in the form of a plan or directive, but a physical change created by the response directive. The physical change corresponds to the change described in the plan and issued by the directive. In the case of the single-response machine, the response function produces the only response the machine is capable of producing.

Logic and Information. Three functions—screening, directive, and response—are characterized by transforming one logical mode into another. The screening mode transforms the raw sensory stimulation into facts about the sensory record (the presence or absence of the ball). The directive function transforms information about the response into the directive or imperative to produce the response. The response function transforms this directive into actual behavior. These various functions are independent of each other, and the content of one cannot be derived logically from the content of any of the others. If they are connected in a performance system, therefore, the connections have to be achieved by invention, not by logic.

The planning function is different from the others in that it links information to information—information about the record (“yes, the ball is present”) to information about the response directive (“yes, it’s appropriate to do response X based on plan X”).

Planning and Directive Functions. It may appear that it would be possible to eliminate the planning step and go directly from the information about the record to the directive. This truncation would simply place two functions on the directive—specifying the response that will be produced and actually directing that response to occur. In practice, this amalgamation of the functions for binary performance systems certainly occurs. From the standpoint of functions, however, the directive function is greatly different from the planning function. The plan is information about what to do. The directive is an imperative to do it.

The necessity of the directive function is not readily identified because the result appears to be automatic. The presentation of the ball simply activates the circuit that produces the response. The role of the directive is demonstrated by wiring the system so that the directive function is simply an on–off switch for the response. If the reception is identified as a positive example, the directive to respond would simply turn on the switch, thereby causing the specified response to be performed.

The function is now observed by disabling the switch but maintaining the operating status for the other functions. The machine receives the sensory input, screens it, identifies the ball, and plans the response to be produced. However, the response does not occur although the system has the capacity to produce the response. The directive to respond has been disabled so there is no basis for activating the response.

The role of the directive function can be demonstrated in another way. Follow this plan or do not follow it: Touch your head.

The plan was the same regardless of whether you followed it. Your system understood what behavior the directions specified. The behavior occurred, however, only if you directed one of your arms to touch your head. “Touch your head,” therefore, was the planning function that described the behavior. What you actually did was the directive function. You either issued the directive “Touch head” or you did not.

Fixed-Response Performance

The planning function becomes more obvious and complex if we add a second behavior—moving the leg in a particular arc. If either “move the leg” or “move the arm” is to be produced in response to the ball on a particular occasion, the planning function of the machine necessarily changes. The identification that the ball is present still activates the planning function, but the planning must provide some form of procedure for making decisions about whether to move the arm or leg.

A large number of decision formats are possible, but the choices fall into four groups: preset sequence, probability, correlational, or combination. For each of these formats, the screening of the positive example would activate the selection process.

Preset Sequence-Based Decisions. The preset sequence would provide a pattern of responses that would be produced on successive trials—simple alternation, double alternation (arm, arm, leg, leg), asymmetrical series (arm, arm, leg, arm, arm), and so on. For the machine to perform any preset series, the machine would need the addition of what amounts to a memory function. It could be as simple as a toggle switch or counter, but it serves a memory function if it carries information about the past into the current setting.

Probability-Based Decisions. Some form of random-selection mechanism would be part of the planning. The system would randomly select on the basis of a schedule that provided for a given probability of arm or leg being chosen on a particular trial. The probability of arm or leg could be weighted so that a given response (e.g., the leg) would be selected on two thirds of the trials. If the system had any type of random selection mechanism, it would be ...

Table of contents

• Cover
• Half Title
• TitlePage
• Copyright
• Contents
• Preface
• Part I: Performance of Nonlearning Systems
• Part II: Basic Learning
• Part III: Extended Learning
• Part IV: Human Learning and Instruction
• References
• Author Index
• Subject Index
• About the Authors