Distribution and Association: Modeling Two Fundamental Principles in Cognitive Control

Cognitive control is pervasive in human behavior. No matter what task one is performing, control of the sequence of one’s actions seems indispensable to accomplish the task. Therefore, control is one of the key issues in understanding human behavior. Moreover, knowledge about how control is exerted in humans may potentially help in designing improved artificial systems, since the control performance humans exhibit is currently unrivaled by artificial systems. Both understanding human control behavior and building improved artificial systems could be achieved by a sufficiently precise computational model of control which, however, currently does not exist. This paper devises a new computational model. In doing so, two fundamental principles of cognitive control in humans are identified. Based on these insights a new model of control which implements those principles is developed. Introduction In everyday behavior1 for every action one takes there are infinitely many other actions possible which one ignores. Instead of reading this text, for instance, you could read something else, contemplate your own research ideas or do anything else you can imagine. Consequently, the question arises ”What makes one take the action one takes and not one of the other possible actions?” or, in other words, ”How is one’s behavior controlled?”. To find an answer to this question is of major importance to psychology, since the issue of control is at the heart of not only some special tasks, but of all human behavior. Moreover, answering this question is also of interest to the field of artificial intelligence, because the control performance of humans exhibits an efficient combination of goal directedness and simultaneous flexibility unrivaled by current artificial cognitive agents/systems. Thus, understanding how human behavior is controlled may help in building better artificial systems. Accordingly, there has been considerable effort to both model control (e.g., Kieras, Meyer, Ballas, & Lauber, 2000; Kimberg & Farah, 2000; Meiran, 2000) and to answer the question of how control is realized in humans (e.g., Allport, Styles, & Hsieh, 1994; Baddeley, 2002; Gopher, 1996; Goschke, 2004; Kluwe, 2000; Norman & Shallice, 1986). An ad hoc answer would be that one’s will causes the action to happen. On a closer look, however, there are several problems with this account of control. First of all, assuming that will controls behavior raises the question who Here and in the following the terms behavior and action will be used in their broadest sense, i.e., including also purely mental behavior/action. or what controls will. Thus, introducing something like will only shifts the problem, leading, in the end, to an infinite regress. A second reason to question that control is exerted solely by will is the fact that in certain situations behavior is seemingly controlled by other factors. For instance, in one classical experimental paradigm, the antisaccade task, participants have to fixate the middle of a computer screen until either on the left or right part of the monitor some stimulus (e.g., a cross) appears. Participants are instructed to look, as soon as the stimulus is presented, at the part of the monitor where the stimulus is not shown, i.e., away from the stimulus. As, among others, Kane, Bleckley, Conway, and Engle (2001) have shown, participants despite their will and effort to look away from the stimulus quite often look towards it. This clearly indicates that there have to be additional determinants of control besides will. As a result, during the last century researchers have tried to clarify the functioning of control of behavior and to explain it without referring to some homunculus. These efforts notwithstanding, Hommel, Daum, and Kluwe (2004) point out that most of how control is exerted is still unknown. This is, as Monsell and Driver (2000) remark, in part due to the lack of a sufficient computational model of control processes which could potentially both constitute a theoretical framework to integrate experimental results and provide a possibility to test different theoretical accounts. Since, as already mentioned, such a model would also be of advantage to the design of artificial cognitive systems its development seems a worthwhile endeavor. Consequently, the conceptualization and implementation of such a model is the aim of the research presented in this paper. To this end, psychological work concerning control is shortly summarized and two fundamental results emerging from this work are identified in the following section. Subsequently, previous models of control are critically discussed with respect to the two fundamental results before a new modeling approach is proposed. As a final point, open issues regarding the new approach are discussed. Fundamental Results From Psychology Despite its importance, control has become a major research topic only comparatively recently, i.e., in the late 1980ies and early 1990ies. Therefore, experimental results regarding control in general which are detailed enough to be used for modeling purposes are virtually nonexistent. Instead, aspects of control have been examined in more depth almost merely with respect to particular activities as, for instance, task switching see Monsell (2003, for a review) or the antisaccade task (cf. Kane et al., 2001). Nonetheless, results emerging from this research point to fundamental principles of how control is realized in humans. Two of these fundamental principles, namely distribution and association, will be described in the following subsections. Control is Distributed It has been an ongoing debate whether control is realized by a unitary component or whether it is the result of many different (probably interacting) components. In neuropsychology, for instance, it has been argued (cf. Norman & Shallice, 1986) that all control is realized by one single brain structure, namely Prefrontal Cortex. Likewise, there are classical psychologists, as Gopher (1996), advocating a single control component. Contradicting to such a unitary component assumption are results from taskswitching experiments (see, e.g., Allport et al., 1994; Monsell, 2003): Certain experimental manipulations do influence task switching performance in a way inconsistent with a unitary component of control. Moreover, evidence from neuroimaging studies, as the one conducted by Garavan, Ross, Li, and Stein (2000), suggest that the Prefrontal Cortex is not the only brain structure involved in control, but that there are many different brain regions distributed over the cortex. Consequently, a majority of researchers (e.g., Allport et al., 1994; Engel, Bertel, & Barkowsky, 2005; Goschke, 2004; Kluwe, 2000) argue that control is the result of the interaction of several different components or, more concisely, that control is distributed. Control is Associative The importance of associations to the problem of control stems from the fact that objects and actions or actions and actions may be associated with each other. Accordingly, stimuli in the environment or the mind may elicit certain actions, i.e., stimuli may induce behavior. This is especially evident every time stimulustriggered actions are inappropriate. One particular example would be looking towards the stimulus in the antisaccade task described above. Although we may be only aware of the influence of stimuli when this leads to error, recent research, as, e.g., Hommel et al. (2004) remark, has shown that associative control is quite pervasive. Unfortunately, however, it is not easy to conceive whether and how association alone enables reasonable behavior in all situations. Considering the large amount of entities that may be associated with one stimulus it is not obvious how such a system of associations could be configured such that it responds correctly in every situation in which this stimulus occurs. In accord with these considerations, psychological work (see Goschke, 2004; Hommel, Ridderinkhof, & Theeuwes, 2002; Kieras et al., 2000) indicates that association, although an important factor, is not the only determinant of control. Instead, a generally accepted account which can be traced back to the beginning of the last century (Ach, 1905) assumes that higher level control processes like intention modulate associative control. If, for example, one’s intention is to determine the sum of two digits this intention will lead to a preselection of associations relating digits with the appropriate action (namely adding). Given this preselection the occurrence of two digits will—associatively, and due to the preselection almost exclusively—elicit the action of adding the two digits. Such a theory can still account for the flexibility of human behavior which can quickly adapt to new situations, while at the same time being able to explain the goaldirectedness, commonly observed in human behavior. Yet, the introduction of the additional factor (i.e., intention) destroys the simplicity of the associative approach, since this factor is potentially homunculus like and needs further explanation. Although the latter issue is not completely settled yet, one idea emerging throughout the field is that intentions stem from perceptions (see, e.g., Kluwe, 2000). In summary, there is considerable evidence showing that control is achieved by an interaction of bottomup (associative, stimulusdriven) and topdown (intention) processes lending the flexibility and directedness to human behavior commonly observed. Having identified two fundamental principles, namely distributedness and associativeness, acting on the control of human behavior, in the following two sections the impact of these fundamentals on modeling control is considered. To this end, existing conceptions of control will be evaluated regarding the principles before a new approach is proposed. Previous Conceptions of Control Since control is at the heart of all human behavior, every model of behavior has to (explicitly or implicitly) implement control in some way or the other. In this section the most renowned and successful of these modeling approaches will be examined with respect to their implementation of control. More precisely, the models considered in this section comprise EPIC (Kieras & Meyer, 1997), ACTR (Anderson et al., 2004), and Soar (Newell, 1990) which will be discussed in turn.