به سوی یک مدل محاسباتی مقایسه اجتماعی: برخی از مفاهیم برای معماری شناختی

Abstract We investigate a general cognitive model of group behaviors, based on Festinger’s social comparison theory (SCT), a prominent social psychology theory. We describe two possible implementations of SCT process at an architectural level, on the basis of the Soar cognitive architecture. The first, which seems to follow directly from Festinger’s social comparison theory, treats the SCT process as an uncertainty-resolution method. The second, takes a different approach, in which an SCT process is constantly active, in parallel to any problem solving activity. We present the implementation of these approaches in the Soar cognitive architecture and argue that one is more suitable for modeling crowd behaviors. In previous work, we have shown that SCT covers a variety of pedestrian movement phenomena. In this paper we present the use of the SCT model in generation of imitational behavior in loosely-coupled groups. Based on experiments with human subjects, we show that SCT generates behavior in-tune with human crowd behavior.

1. Introduction Models of crowd behavior facilitate analysis and prediction of the behavior of groups of people, who are in close geographical or mentally similar states, and are affected by each other’s presence and actions. Existing models of crowd behavior, in a variety of fields, leave many open challenges. In social sciences and psychology, models often offer only qualitative description, and do not easily permit algorithmic replication. In computer science, models are often simplistic, and typically not tied to specific cognitive science theories or data. Moreover, existing computer science models often focus only on a specific phenomenon (e.g. flocking, pedestrian movement), and thus must be switched depending on the goals of the simulation. In our previous work (Fridman & Kaminka, 2007), we presented a model of crowd behavior, based on social comparison theory (SCT) (Festinger, 1954), a popular social psychology theory that has been continuously evolving since the 1950s. The key idea in this theory is that humans, lacking objective means to evaluate their state, compare themselves to others that are similar. We believe that social comparison is a general cognitive process underlying the social behavior of each individual in crowd. However, it was described as a stand-alone algorithm, with no discussion of how it should be integrated into the action–selection processes of the agent. Moreover, the model was evaluated almost entirely in the domain of synthetic pedestrian movement, without comparison to human crowd behavior. In this paper we describe the implementation and adaptation of the SCT model in the Soar cognitive architecture, and provide a detailed description of its use in modeling imitational behavior. We describe two implementations of SCT process at an architectural level. The first, which seems to follow directly from Festinger’s social comparison theory, treats the SCT process as an uncertainty-resolution method, i.e., as a weak (read: general) problem-solving method, which is social. The second takes a different approach, in which an SCT process is constantly active, in parallel to any problem-solving activity. We argue that the latter approach, in which comparison is a continuous process, is more suitable for modeling crowd behaviors. In addition, we evaluate the use of SCT in generation of imitational behavior in studies with human subjects. We show that SCT generates behavior in-tune with human crowd behavior: The subjects ranked SCT to be a middle-ground between completely individual behavior, and perfect synchronized (“soldier-like”) behavior. Independently, human subjects gave similar rankings to short clips showing human crowds

6. Conclusions This paper presented a model describing crowd behavior, inspired by Festinger’s social comparison theory (Festinger, 1954). The model intuitively matches many of the characteristic observations made of human crowd behavior, in particular as described by major crowd behavior theories. We presented in detail two implementation approaches to social comparison processes, in Soar cognitive architecture. The first approach treats social comparison as a problem-solving activity, to be triggered by uncertainty in the agents’ task-oriented reasoning. The second approach treats social comparison as an on-going processes, taking place in parallel to any task-oriented reasoning and actions. In early experiments, it was shown that the latter approach is superior, and we discuss the reasons for this choice in depth. We report on experiments with human subjects, to evaluate the use of the on-going SCT process in modeling imitational behavior in a group. Though there is a lack of objective data against which the model can be evaluated, results of experiments with human test subjects are promising and seem to match intuitions as to observed behavior. The subjects ranked SCT to be a middle-ground between completely individual behavior, and perfect synchronized (“soldier-like”) behavior. Independently, human subjects gave similar rankings to a short news clip showing human crowds. The research reported in this paper raises several implications for cognitive architecture and their structure, as it pertains to social reasoning. • First, accepting that social comparison takes place in parallel to problem-solving activity implies that agent modeling – the process by which an agent keeps track of other agents, monitors their behavior, and possibly infers their intention–is carried out at an architectural level, in parallel. Based on this, it would seem that agent modeling, which forms the basis for social comparison, is in some sense like learning: Always present, always taking place. The agent can choose, perhaps, to what degree to follow up on its results, but the agent cannot turn it off. • Second, accepting that social comparison takes place constantly, in parallel to task-oriented reasoning, we are now faced with the challenge of investigating the timing of comparison: When do humans follow up on the results of the comparison, and when do they ignore it? Does the tendency to follow up on social comparison change between people? What affects this tendency? etc. • Third, there is of course much more depth to social comparison, as the basis for cognitive modeling. For instance, we know from psychology literature that the comparison process itself is more complex than the process captured in Algorithm 1. For instance, psychology tells us that the process of selecting a target for comparison is much more involved than simply picking the most similar agent (Suls, 2000). . We hope to pursue these implications and directions for further research in our future work. We are particularly interested in continuing to investigate cognitive models of social behavior and social reasoning.