Article Preview
TopBackground
Norms are defined as either rules of expected behavior in a society or as behavior that is common in a society (Morris-Martin et al., 2019). Norms provide a means to regulate agent behavior, which requires some consideration of how norms impact agent reasoning and behavior (Finnemore & Sikkink, 1998). In open, dynamic societies, agents must work with others who do not necessarily have the same set of objectives. If left unchecked, self-interested agents will try to accomplish their individual goals without regard for others. The concept of the normative life cycle (NLC) (Finnemore & Sikkink, 1998; Frantz & Pigozzi, 2018; Hollander & Wu, 2011; Mahmoud et al., 2014; Savarimuthu & Cranefield, 2011) has been used to explain the mechanics at a conceptual level, involved or should involve in designing norm aware agents. The concept of norm synthesis—the generation of normative rules—is derived from most of the NLC processes (Frantz & Pigozzi, 2018). Specifically in regard to the aspect of generating normative rules at run time, the formalization of the NLC into a workable and practical framework using machine-learning techniques is best described by Morales et al. (2015b) as the IRON framework. Their automated normative rule-generation method requires the designer to explicitly specify the conflict state for the framework to synthesize normative rules to avoid said conflict state via case-based reasoning. An alternative approach that reduces designer-defined conflict states proves beneficial in designing adaptable NorMAS, and one of those approaches is using RL.
Norm Synthesis
Norm synthesis is designing the emergence of applicable normative rules or policies (Shoham & Tennenholtz, 1992). Norm synthesis can be described as the process of generating rules that are not hard-coded recipes presenting reactive behaviors, such as those in static expert systems, but rather describe consequences arising from observations for reasoning about the current context, resulting in situation-specific norms (Lee et al., 2014). Alternatively, in an abstract state transition system, norm synthesis generates rules for avoiding undesirable conflict states, leading to conflict-free states (Christelis & Rovatsos, 2009) while also aiming at developing a stable set of norms (Morales et al., 2015a). Adopting the viewpoint of Frantz & Pigozzi (2018), we interpret the concept of norm synthesis as an important norm identification/detection approach, which is further supported by the viewpoint that norm emergence can be used as a source for norm synthesis in NorMAS (Morris-Martin et al., 2019). We believe that the current understanding of norm synthesis constitutes the cyclic process of norm emergence (or convention emergence) and norm detection and representation, as shown in Fig. 1. The goal of norm synthesis, then, lies in prescribing and converting implicit norms (social) into explicit norms (legal) represented in a structured format, i.e., rule-based or logical representation (Morris-Martin et al., 2019), making the transfer and interpretability of learned norms possible.
Figure 1. Current Understanding of Norm-Synthesis Framework