Fostering Multi-Agent Cooperation through Implicit Responsibility

For integration in real-world environments, it is critical that autonomousagents are capable of behaving responsibly while working alongside humans andother agents. Existing frameworks of responsibility for multi-agent systems typically model responsibilities in terms of adherence to explicit standards. Suchframeworks do not reflect the often unstated, or implicit, way in which responsibilities can operate in the real world. We introduce the notion of implicit responsibilities: self-imposed standards of responsible behaviour that emerge and guideindividual decision-making without any formal or explicit agreement.We propose that incorporating implicit responsibilities into multi-agent learningand decision-making is a novel approach for fostering mutually beneficial cooperative behaviours. As a preliminary investigation, we present a proof-of-conceptapproach for integrating implicit responsibility into independent reinforcementlearning agents through reward shaping. We evaluate our approach through simulation experiments in an environment characterised by conflicting individual andgroup incentives. Our findings suggest that societies of agents modelling implicitresponsibilities can learn to cooperate more quickly, and achieve greater returnscompared to baseline.