Evaluating computational models of ethics for autonomous decision making

Abstract

Computational models for ethical autonomy, are crucial for building trustworthy autonomous systems. While different paradigms of ethical autonomy are pursued, comparing and contrasting these paradigms remains a challenge. In this work, we present SPECTRA (Strategic Protocol Evaluation and Configuration Testbed for Responsible Autonomy) a general purpose multi-agent, message-passing framework on top of which, different models of computational ethics can be implemented. The paper also presents our implementation of four paradigms of ethics on this framework– deontology, utilitarianism, virtue ethics and a recently proposed paradigm called computational transcendence. We observe that although agents have the same goal, differences in their underlying paradigm of ethics have a significant impact on the outcomes for individual agents as well as on the system as a whole. We also simulate a mixed population of agents following different paradigms of ethics and study the emergent properties of the system.