Motives and Emotions in a General Learning System

Introduction

The relationship between motives, emotions and learning in the design of intelligent systems has received relatively little consid­ eration to date. Sloman and Croucher (1981) include mechanisms for changing the current set of motives and the need to learn motives in their list of constraints on the design of intelligent systems which determine what would make them well adapted. They, also, point out that a complex and changing collection of motives and the emotions they generate must have a profound effect on what is learned when and contribute to enormous variations between individuals. Pfeiffer (1982) in presenting the general framework of a theory of emotion in cognitive science terms lists learning processes and evolutionary considerat­ ions among the important issues which have not yet been addressed. His theory has been partially specified in a computer simulation model called FEELER which is implemented in an extended version of PRISM (Langley and Neches, 1981).

Clarification of the nature of the relationships between motives, emotions and learning in intelligent systems involves tackling an agenda of fundamental items which can be conveniently described in terms of human development. It is necessary to explore the survival value or adaptive function of motives and emotions in intelligent systems from an evolutionary perspective. What motivational and emotional processes must be assumed to be wired in as features of innate

endowment? How does the learning of additional

MOTIVES AND EMOTIONS IN A GENERAL LEARNING SYSTEM

J.G . Wallace

Centre for Research on Intelligent Systems Deakin University, Geelong, Vic., 3217. Australia.

motives and emotions proceed as a response to environmental interaction? What are the results of the influence of motives and emotions on learning processes and, conversely, the effect of learning processes on the nature of motives and emotions? The remainder of this paper reports a preliminary exploration of these issues employing BAIRN, a general learning system, and adopting a perspective on motives and emotions derived from the work of Sloman and Croucher and Pfeiffer.

BAIRN: A General Learning System

As a comprehensive account of BAIRN has been provided elsewhere (Wallace, 1982) only the features directly involved in the construction and representation of motives and emotions will be described. The BAIRN project is aimed at the construction of a theory of cognitive development in the form of a self­ modifying information processing system in which productions are employed as the units of cognitive analysis. The specific objective is the construction of a computer program as a completely explicit and demonstrably sufficient version of the theory.

The BAIRN system comprises general learning processes and a world model which governs the nature of its interaction with theenvironment. An integrated form of representation for procedural and declarative knowledge is adopted in the long term memory (LTM) world model. It is composed of a network made up of nodes at which are stored production systems. The mode of operation of the network is based on three levels of consciousness (focal, peripheral and unconscious) ~hich permit a limited number of nodes to be active in parallel. Each node in the network has prepotency values which define the strength of its bid to be allowed to operate when specific inputs suitable for its activation are present. This feature provides an opportunity to explore the effects of variations in motivational structure on the system's operation.

The approach adopted in the BAIRN project maximises emphasis on general learning procedures. The learning processes are responsible for the creation and addition of new productions to LTM. Discharging this function requires the ability to monitor the results of environ­ mental interaction. The mechanism adopted is a time line or episodic memory which contains a sequential record of the system's activity. If any regularities exist in interaction with the environment they are represented in the time line . Self-modification takes place through the detection of this regularity by the learning processes and the subsequent addition to LTM of productions which will capitalize upon it . Regularity detection involves two probabilistic features which constitute dimensions of individual difference. The first is the M(atching) I(ndex) C(riterion) which has a constant value for each individual. The MIC determines the level of agreement before two sequences are regarded as a match. The second probabilistic feature is the C(ritical) C(onsistency) L( evel). Unlike the MIC, the CCL exhibits variations within a band characteristic of the individual. It reflects both the number of novel sequences and the degree of agreement between them which comprise a level of consistency warranting modification of the LTM world model.

Adaptive Function of Motives and Emotions

In information processing terms the existence of motives can be attributed to the need for determination of directionality in intelligent systems endowed with a high degree of flexibility or adaptability. Appropriate constraints on the wide range of possible courses of action are essential for the establishment and maintenance of successful environmental interaction. The first step towards fulfilling this function in BAIRN is the addition of motive generators to the system's initial repertoire of processes. Motive generators bias the construction of the system's representation of its world in particular directions and, as a result, constrain and direct its behaviour. They operate by scanning the time line trace for sequences representing experience falling within the area of their motives and initiating appropriate additions or modifications to LTM . This process is conducted both on-line and off-1 ine.

Emotions can be regarded as 'hard wired' sources of information that assist in a continuing process of discrimination which is crucial in canalising the behaviour of the system in appropriate directions. Episodic records of physiological activity associated with emotions assist motive generators in performing their function by facilitating rapid detection and discrimination of time line segments both on-line and off-line. These records initially reflect the operation of primary emotion generators included in the system's 'innate' repertoire. These are productions linking sensori-perceptual conditions with physiological actions defining primary emotions such as pleasure, fear, anger and so on. The physiological trace data is detected and reacted to by 'innate' motive generators. Their conditions include patterns of physiological activity defining primary emotions, indications of the level of emotional intensity derived from the intensity of sensory stimulation, and other situational features defining the area of the motive . In innate motive generators these situational features are of a relatively global nature, such as the perception of human contact.

. T.he. assignment .of distinct, physiological def1n1t1ons to emot1ons via primary emotion generators is at variance with the approach adopted in FEELER. Since a general leve l of autonomic arousal is assumed the physiological component in FEELER cannot be used to distinguish between different emotions. This objective is achieved by means of distinctions in the cognitive-evaluative aspects of emotions. In addition to employing specific variations in p~ysiological arousal the present approach reta1ns a fundamental function for the general level of autonomic arousal. It is assumed that it contributes to determining the nature of the spread of activation that occurs in the LTM network during each cycle. High arousal produces the lowest level of contextual inhibition and the least constrained spread. Low arousal is associated with the most constrained spread and contributes to 'concentrated attention' while intermediate arousal leads to a controlled spread of activation well suited to complex processing such as problem solving.

In~eraction between the initial repertoires of emo~1on generators and motive generators determ1nes the episodic records which are selected as starting points for off-line processing of the time line in the search for consistent sequences . In some cases mot~ve ge~erators representing highly significant m~t1ves d1rectly determine that a new node w1ll be added to LTM on the basis of a single occurrence of a sequence. In others the decision depends on the result of time line processing by the consistency detection proc:sses employing the MIC and CCL parameters :ar:1ed out after a motive generator has 1nd1cated t~e.starting point for analysis . When an add1t1on to LTM occurs an initial prepotency value is assigned to the new node by th7 motive generator(s) involved in its creat1on. The prepotency reflects the status of. the moti~e generator(s) in a hierarchy of pr1mary mot1ves and specific values of the emotional intensity and situational feature variables included in the sequence.

.since records of the operation of primary e~o~1on ~enerators are included in sequences g:v1ng r1se to new nodes, productions providing ~1rect access to the generators are incorporated 1n.the definition lists of the new nodes. Th1s en~ures that future activation of the nodes w1ll be accompanied by experience of the associated emotions.

Assigning of prepotencies is not confined to newly created nodes. Individual nodes c~pare the results of their activations ~1try exp~ctations based on previous input/output on]unct1ons stored on their description lists. Unexpected or novel outcomes are flagged in the time line record to enable processing by motive generators and assignment of a prepotency value. The effect of this procedure is to provide each node with a range of context linked prepotencies as environmental interaction proceeds. This replaces the initial global prepotency.

General functional characteristics desirable for effective environmental interaction and facilitation of learning can be conferred on the system by additions to the initial repertoire of motive generators. Frequent activation of newly constructed nodes encourages detection of their relationships with other nodes and contributes to further modification and enrichment of the LTM world model. This effect can be produced by a motive generator which temporarily increases the initial prepotency level of new nodes on a decrement with repetition basis.

A similar approach can achieve the selective attention to new environmental features characteristic of human performance. In BAIRN repeated activation of nodes within a parameterised range of cycles is indicated by markers on the time line records of the repetitions. If 'newness' is defined as the absence of repetition markers, a motive generator monitoring time line entries online can produce a focussing of attention by temporarily increasing the prepotency level of the nodes responsible for the 'new' time line records. Such increments would disappear with fewer repetitions of activation than the temporary increases accorded to new nodes.

Learning of Emotions and Motives

Interaction of the 'innate' repertoires of emotion and motive generators not only contributes to construction of the system's world model but results in the learning of further emotions and motives. Representations of higher level or compound emotions are derived from time line sequences containing activations of two or more primary emotion generators or two or more nodes with generators featuring on their definition lists. Higher level motive generators capable of processing the more complex physiological trace data produced by the compound emotion generators are derived from time line sequences including associated activations of the initial motive generators which process primary emotions. These compound motive generators provide a capability to assign initial prepotency values to new nodes derived from sequences featuring compound emotions.

Interaction of Emotions, Motives and Learning Processes

BAIRN provides illustrations of the effect of learning processes on emotions and motives as well as the influence of emotions and motives on learning processes. The settings of the MIC and CCL parameters determine the generalisation/discrimination boundary Wundy, 1982) that operates in the construction of new nodes. As a result wide individual differences in the specifics of the representations of compound emotions and motives constructed by a system will be produced by varying the combinations of MIC and CCL values involved in detecting the sequences from which they are derived. Combinations emphasising the extremes of generalisation or discrimination will resul"t in emotions and motives which in humans are associated with pathology.

This interaction offers the prospect of endowing self-modifying intelligent systems with the basis of distinctive 'personalities' that will emerge as construction of their world model proceeds. This would be achieved by manipulation of the initial repertoires of emotion and motive generators and the values of the learning parameters. The possible membership of the "generator repertoires is not restricted to the range of human emotions and motives. It is defined by the range of features of intelligent systems which are functionally equivalent to emotions and motives. The objective of providing intelligent systems with emerging personalities is to allow for self-modification within broad constraints. The system retains the flexibility and general learning capacity necessary to construct a world model and maximise the effectiveness of interaction on the basis of the specifics provided by the environment. The environment, however, does not dictate the overall emphasis in the system's operation. Systems designed 6n this basis would be suitable for the control of third generation robotic devices required to function for protracted periods without human supervision.