Norm Avoidance and Reinforcement Learning: Definitions and Analys

Abstract

Recent work in the field of normative reinforcement learning has proposed various methods for teaching agents normative behaviors. However, these approaches often face the problem of norm avoidance, where the learning agent follows the formal rules of a norm but, without inciting a direct violation, circumvents its intended purpose by exploiting loopholes to achieve its own goals. In this paper, we present a range of approaches to address the problem of norm avoidance, as well as benchmark examples. We then compare these approaches by analysing their behavioral discrepancies in relation to the benchmarks. Finally, we discuss how each approach mitigates norm avoidance, and we pinpoint the constraints inherent to each approach.