Social Intelligence as Norm Adaptation

Machine learning is the core of artificial intelligence. Letting machine learning be driven not only by individual goals, but by goals consistent with those of a coalition of agents, is very difficult. Successful inductive rules for agent behaviour are typically based on machine guesses, the quality of which are measured in terms of precise real numbers representing utility. Random guesses are of low quality, guesses based on heuristics are of the same quality as the heuristics themselves, and guesses based on entire theories are of the quality of the theories together with the quality of the rules for action and their relation to the theory. Such underlying theories have in machine learning thus far not been theories for social action, but theories for individual action. Analogously, the effects of actions on other agents have been studied mainly as feedback to the agent at hand, ignoring the modelling of the utility assessments of those other agents. We propose that intelligent agent action be studied with respect to a social space. The latter consists of a number of agents, their assessments, as well as their sets of norms. Norms are here treated technically, as constraints on individual action. The learning of new norms, and the strife of each agent to act in keeping with the norms of the coalitions of which it is a member constitutes social intelligence.