Socially Intelligent Technology

Major improvements in the ability of artificial agents to deal with humans and to emulate humans will require those artifacts to be socially intelligent. Socially intelligent agents could serve as intelligent tutors, nannies, personal shoppers, etc. Sets of socially intelligent agents could be used to emulate human groups/organizations to determine the relative efficacy, feasibility, and impact of new technologies, legislation, change in policies, or organizational strategy. At issue are questions of how social these agents need to be and what is the basis for socialness. It is relatively easy to create artificial agents that are more capable than a human for a specific well-understood task. It is relatively easy to create artificial agents that can, in a limited domain, act like humans. But these factors do not make the agents generally socially intelligent. One of the research challenges will be for computer scientists and social scientists to work together to develop artificial social agents. Such agents should be social at both the cognitive and precognitive (bio) level. Current approaches here are software-limited. They are also potentially limited by data; nanotechnology, which will enable higher levels of storage and processing, will also be necessary. That is, creating large numbers of cognitively and socially realistic agents is technically unfeasible using a single current machine. Yet, such agents need to exist on a single machine if we are to use such tools to help individuals manage change.

A key component of socialness is the ability to operate in a multiagent environment (Epstein and Axtell 1997; Weiss 1999). However, not all multiagent systems are composed of socially intelligent agents. For a machine to be socially intelligent, it needs to be able to have a "mental" model of others, a rich and detailed knowledge of realtime interaction, goals, history, and culture (Carley and Newell 1994). Socially intelligent agents need transactive memory, i.e., knowledge of who knows whom (the social network), who knows what (the knowledge network), and who is doing what (the assignment network). Of course this memory need not be accurate. For agents, part of the "socialness" also comes from being limited cognitively. That is, omniscient agents have no need to be social, whereas, as agents become limited — boundedly rational, emotional, and with a specific cognitive architecture — they become more social.

One of the key challenges in designing machines that could have such capabilities is determining whether such machines are more or less effective if they make errors like humans do. What aspects of the constraints on human cognition, such as the way humans respond to interrupts, the impact of emotions on performance, and so on, are critical to acquiring and acting on social knowledge? While we often see constraints on human cognition as limitations, it may be that socialness itself derives from these limitations and that such socialness has coordinative and knowledge benefits that transcend the limitations. In this case, apparent limits in individuals could actually lead to a group being more effective than it would be if it were composed of more perfected individual agents (Carley and Newell 1994).

A second key challenge is rapid development. Computational architectures are needed that support the rapid development of societies of socially intelligent agents. Current multiagent platforms are not sufficient, as they often assume large numbers of cognitively simple agents operating in a physical grid space as opposed to complex intelligent, adaptive, learning agents with vast quantities of social knowledge operating in social networks, organizations, and social space. Moreover, such platforms need to be extended to enable the co-evolution of social intelligence at the individual, group, and organizational level at differing rates and accounting for standard human processes such as birth, death, turnover, and migration.

A third challenge is integrating such systems, possibly in real time, with the vast quantities of data available for validating and calibrating these models. For example, how can cities of socially intelligent agents be created that are demographically accurate, given census data?

0 0

Post a comment