Organizational design principles and techniques for decision-theoretic agents

Recent research has shown how an organization can influence a decision-theoretic agent by replacing one or more of its model components (transition/reward functions, action/state spaces, etc.), and how each of these influences impacts the agent’s decision-making performance. This paper delves more precisely into exactly which parts of an agent’s model should be organizationally influenced, and asserts a broader principle for delineating what aspects of an agent’s behavior an organization should be sanctioned to influence. We present a formal framework for specifying factored organizational influences and incorporating them into agents’ decision models, and empirically demonstrate that organizational specifications based on our proposed principle outperform the alternatives. We further describe an algorithm for automating the organizational-design process that is inspired by this principle, and demonstrate empirically that its organizational designs are both intuitively sensible and also find and exploit domain structure that our hand-generated designs miss.