Emergent Responsible Autonomy in Multi-Agent Systems

Autonomous agents operating in multi-agent environments, face the dilemma of responsibility where they must choose between actions that are individually beneficial versus those that are considered responsible and ethical. Current approaches address this problem either using external reinforcements, or intrinsic notions of ethics that act as constraints overriding the agents' rational choice. Both of these approaches become difficult to scale across complex, dynamic situations, where the responsibility dilemma has to be resolved dynamically. Thus, there is a need to design models of agency, where a sense of ethics and responsibility are an integral part of the agent model and not in conflict with agents' self-interest dynamics. Towards this end, this thesis proposes a model called Computational Transcendence (CT) in which, an agent's "sense of self" is made elastic, that enables it to dynamically identify with external elements of its environment like other agents, communities and concepts. Agents continue to act rationally, working towards utility maximisation, but their utility is determined by their sense of self formed from their elastic identities. We show that this leads to emergence of responsible autonomy, in various multi-agent network conditions.