A Computational Architecture for Social Agents

This article describes a new class of information-processing models for social agents. They are derived from primate brain architecture, the processing in brain regions, the interactions among brain regions, and the social behavior of primates. 1. Primate brain and behavior In another paper (Bond, 1996), we have reviewed the neuroanatomical connections and functional involvements of cortical regions. We reviewed the evidence for a hierarchical architecture in the primate brain. By examining neuroanatomical evidence for connections among neural areas, we were able to establish anatomical regions and connections. We then examined evidence for specific functional involvements of the different neural areas and found some support for hierarchical functioning, not only for the perception hierarchies but also for the planning and action hierarchy in the frontal lobes. We concluded that the overall structure, both anatomical and functional, of the primate brain is that of a set of cortical regions with spatial localization and clustered interconnectivity. Each region is specialized in data processing and data storage to a limited number of data types. Regions process data in about 20 milliseconds, and communication channels comprise about 1 million fibres each carrying a firing rate, perhaps usable as 8 bits of information. Cortical regions are connected as a perception hierarchy, a planning and action hierarchy, and connections between corresponding levels of these two hierarchies. The perception hierarchy is based on data representing increasingly general situations. The planning and action hierarchy is based on increasingly general situations, plans and control. Hierarchical organization as an organizing principle for brain architecture has been pointed out by Albus (1981). Hierarchical control architectures have been discussed by Mesarovic et. al. (1970). From a study of primate behavior, as reported for example by Cheney and Seyfarth (1990). we concluded that affiliative behavior is fundamental, and that joint plans are co-constructed (Goodwin 1981) by primates interacting socially. Humans have similar behaviors (Harcourt and De Waal 1992). It seems that most of monkey behavior can be categorized under a small number of headings, and that almost all this behavior is socially interactive. That is, behavior of any individual monkey is highly coordinated with and motivated by that of other individuals, and involves detailed interaction with others. The following list covers most types of primate behaviors: (i) grooming, (ii) mating, (iii) infant caretaking, (iv) dominance struggles, (v) foraging and feeding, and (vi) territorial definition and control. Most of these are not solitary actions but involve detailed interaction with others. Even apparent solitary action is in any case strongly and continuously conditioned by the presence of others. We should naturally view these behaviors as joint behaviors carried out by two or more monkeys. Figure 1 shows a vervet monkey. Figure 1. A vervet monkey 2. Our computational architecture To obtain a computational architecture for a social agent based on the primate brain, we take the functional hierarchy composed of neural regions, and interpret each region as a computational module. Each module will have processing and storage abilities. We take connections between neural regions to be communication channels that data may be transmitted along. Figure 2 diagrams our social agent model. The top of the hierarchy is towards the front (left in the diagram), with perception below and action above.