Modeling Interaction Between Metacognition and Emotion in a Cognitive Architecture

While research in metacognition has grown significantly in the past 10 years, there has been a relative lack of research devoted to the focused study of the interactions between metacognition and affective processes. Computational models represent a useful tool which can help remedy this situation by constructing causal models of demonstrated correlational relationships, and by generating empirical hypotheses which can be verified experimentally. In this paper we describe enhancements to an existing cognitive–affective architecture that will enable it to perform a subset of metacognitive functions. We focus on modeling the role of a specific metacognitive factor, the feeling of confidence (FOC), and the anxiety-linked metacognitive strategy of emotionfocused coping. Introduction and Objectives Metacognition, defined variously as ‘thoughts about thoughts’ or ‘awareness and control’ of one’s thoughts, is considered by many to be an essential component of skilled performance, influencing memory functions, learning and skill acquisition, and problem-solving. While research in metacognition has grown significantly in the past 10 years, there has been a relative paucity of research devoted to the focused study of the interactions between metacognition and affective processes. A notable exception is the work of Wells (2000), and Matthews and Wells (2004). However, this work focuses on metacognition and emotion in the context of emotional disorders (e.g., depression, generalized anxiety, obsessive compulsive disorders), rather than on the role of emotion in normal subjects, or the interaction between emotion and metacognition in transient affective states (e.g., high stress or frustration). Computational models represent a useful tool for elucidating the mechanisms of cognitive processing. Ideally, these models are based on existing empirical data and used to confirm hypothesized mechanisms. However, in situations where such data are lacking, these models can be helpful in generating empirical hypotheses for further experimental testing and data gathering. In this paper we describe enhancements to an existing cognitive–affective architecture (MAMID) (Hudlicka, 2002; 2003), which support the exploration of affectivemetacognitive interactions. The enhancements will enable MAMID to perform a subset of the metacognitive functions involved in monitoring and control of cognition, and the associated metacognitive knowledge and beliefs. The initial focus will be on modeling the role of a specific metacognitive factor, the feeling of confidence (FOC), and the anxiety-linked metacognitive strategy of emotionfocused coping. The intended benefits of the model are in both the theoretical and the applied realms. In the theoretical realm, the exercise of building a model requires an operationalization of concepts and relationships which help refine existing psychological theories, and generate empirical hypotheses for further testing. In the applied realm, the explicit model of metacognition, and its interactions with affective factors, promises to provide a more realistic model of human behavior, both adaptive and maladaptive (e.g., models where metacognition diminishes performance), and generate more effective agent behavior (e.g., improved performance under stress). This paper is organized as follows. First, we briefly summarize key findings about metacognition and about known interactions between metacognition and emotion. Next, we describe the existing MAMID cognitiveaffective architecture, within which the proposed metacognitive component will be implemented. We then describe the proposed design for modeling the role of FOC in metacognitive monitoring and control, and the metacognitive knowledge required. We then outline two models where emotion influences metacognitive processing. We conclude with a brief discussion of related work, a summary and future work. In Proceedings of the AAAI Spring Symposium on Metacognition in Computation. AAAI Technical Report SS-05-04. Menlo Park, CA: AAAI Press. pp. 55-61. (2005) 2 Metacognition and Its Interaction with Emotion What is metacognition? The simplest definition of metacognition is “thinking about thinking” (Nelson, 2002). However, the simplicity of this statement belies the complexity and diversity of processes and structures that mediate the variety of identified metacognitive activities. More differentiated definitions distinguish between awareness (and associated monitoring functions of cognition), and control (and associated executive and selfregulatory functions of cognition) (Osman and Hannafin, 1992; Nelson and Narens, 1990). A more encompassing definition states that metacognition is a “multifaceted concept comprising the knowledge and beliefs, processes and strategies that appraise, monitor or control cognition” (Wells, 2000). Metacognitive knowledge is then defined as knowledge individuals have about their own cognitions, as well as about the factors that influence their cognitions. What is the role of metacognition? Evidence indicates that metacognitive control and regulation is comprised of a range of functions including attention allocation, checking, planning, memory retrieval and encoding strategies, and detection of performance errors (Wells, 2000). In general, metacognition is involved in strategy selection for complex problems requiring resource tradeoffs, for dealing with unfamiliar situations, and for troubleshooting. A number of researchers discuss the fact that metacognition can be helpful, neutral, or harmful to cognition and performance (e.g., Paris, 2002). Relationship Between Metacognition and Emotion As stated above, data regarding the mutual influences among emotion and metacognition are unfortunately limited, and focused almost exclusively on psychopathology (e.g., Wells, 2000; Matthews and Wells, 2004). For the purpose of modeling, we need to identify the specific effects of particular affective factors (states or traits) on particular metacognitive functions and knowledge. To help organize the known effects, and to identify gaps in data, it is useful to categorize the effects into those resulting from states vs. traits, and those affecting processing mechanisms vs. knowledge structures. Examples of identified correlations include: State effects on processes: Anxiety-linked appraisal of events as threats; emotion-focused coping; Depression-linked self-criticism focused coping; Trait effects on processes: Neuroticism-linked preference for self-information; Trait effects on knowledge: Neuroticism-linked predominance of negative schemas (threat, negative self evaluations, negative future projections). MAMID Cognitive-Affective Architecture and Modeling Methodology Here we briefly describe the existing cognitive-affective architecture which will be enhanced with the proposed metacognitive functions. A key component of the architecture is an affect appraisal module, which dynamically generates affective states as a function of both internal and external factors (e.g., incoming cues, internal situation assessments and goals), and both dynamic and static agent attributes (e.g., prior existing emotion, stable personality trait profile). We also discuss the generic modeling methodology used to model the interacting effects of states, traits and other individual differences in terms of parametric manipulations of the architecture processes and structures. MAMID Cognitive Architecture The cognitive architecture implements a sequential see-think-do processing sequence (figure 1), consisting of the following modules: sensory pre-processing, translating incoming data into task-relevant cues; attention, filtering incoming cues and selecting a subset for processing; situation assessment, integrating individual cues into an overall situation assessment; expectation generation, projecting current situation onto possible future states; affect appraiser, deriving the affective state (both valence and four of the basic emotions) from a variety of external and internal elicitors, both static and dynamic; goal selection, selecting critical goals for achievement; and action selection, selecting the best actions for goal achievement. These modules map the incoming stimuli (cues) onto the outgoing behavior (actions), via a series of intermediate internal representational structures (situations, expectations, and goals), collectively termed mental constructs. This mapping is enabled by long-term memories (LTM) associated with each module, represented in terms of belief nets or rules. Mental constructs are characterized by their attributes (e.g., familiarity, novelty, salience, threat level, valence, etc.), which influence their processing; that is, their rank and the consequent likelihood of being processed within a given architecture cycle; (e.g., cue will be attended, situation derived, goal or action selected). (Note that the availability of the mental constructs from previous frames of the execution cycle allows for dynamic feedback among constructs, and thus departs from a strictly sequential processing sequence.) In Proceedings of the AAAI Spring Symposium on Metacognition in Computation. AAAI Technical Report SS-05-04. Menlo Park, CA: AAAI Press. pp. 55-61. (2005) 3 Figure 1: MAMID Cognitive Architecture: Modules & Mental Constructs The Affect Appraisal module is a core component of the MAMID architecture. It integrates external data (cues), internal interpretations (situations, expectation) and priorities (goals), and stable and transient individual characteristics (traits and existing emotional states), and generates an affective appraisal in terms of both a valence (positive / negative) and one of the four basic emotions (fear/anxiety, anger/frustration, sadness, joy). The basic emotions are calculated via a series of belief nets stored in the agent’s LTM. Differences in the triggering elicitors for particular emotions allow for the representation of individual idiosyncracies in emotion triggering (e.g., Agent A might react to situation x with anger, Agent B with fear, whereas Agent C might not have an affective reaction at all.) The model incorporates elements of several recent appraisal theories: multiplelevels and multiple stages (Leventhal & Scherer, 1987; Smith & Kirby, 2001). Generic State / Trait Modeling Methodology To model the interacting effects of traits and states on cognitive processing, MAMID uses a previously described methodology (Hudlicka, 2002; 1998), which consists of mapping particular state / trait profiles onto specific architecture parameter values (figure 2). These parameters then control processing within individual architecture modules. Functions implementing these mappings were constructed on the basis of the available empirical data. For example, reduced attentional and working memory (WM) capacity, associated with anxiety and fear, are modeled by dynamically reducing the attentional and WM capacity of the architecture modules, which then reduces the number of constructs processed (fewer cues attended, situations derived, expectations generated, etc.). Attentional threat bias is modeled by higher ranking of threatening cues, thus increasing their likelihood of being attended, and by higher ranking of threatening situations and expectations, thus increasing the chances of a threatening situation / expectation being derived. Traitlinked structural differences in LTM are supported by allowing the flexible selection of alternative LTM clusters, reflecting distinct personality traits. Traits also influence the dynamic characteristics of the emotional responses (ramp up, decay, and maximum intensities). Figure 2: Parametric State / Trait Modeling Methodology The initial version of MAMID was implemented in the context of a peacekeeping scenario, with each instance of the MAMID architecture controlling the behavior of a simulated Army commander, reacting to a series of surprise situations (e.g., ambush, hostile crowd) (Hudlicka, 2003). MAMID was able to demonstrate distinct processing differences due to the different trait profiles and dynamically generated states, with the distinct commanders behaving differently during the course of the scenario, leading to differences in mission outcomes. The domain-independent MAMID architecture is currently being transitioned into a NASA context, and it is within this context that the metacognitive enhancements described below will be implemented. Proposed MAMID Enhancements Implementing Metacognitve Functions The objective of the enhancements described below is to augment the existing MAMID architecture with the ability to perform a subset of metacognitive functions, and to explicitly model interactions among selected metacognitive functions and emotion. Below we describe a model of the feeling of confidence factor, and its role in triggering metacognitive control strategies. In the next section we describe two examples focusing on emotion. Feeling of Confidence (FOC) FOC is a component of metacognition that reflects the level of confidence in particular cognitions. FOC judgments can refer to past, present and future cognitive activities, and apply to a variety of cognitive processes, including memory retrieval, problem-solving, and planning. FOC is thought to play a role in controlling cognitive iteration during problem-solving and memory retrieval, by determining whether a particular answer will be accepted, or whether Cues