Modelling Interactions via Commitments and Expectations

Introduction

Organizations can be seen as sets of entities regulated by mechanisms of social order and created by more or less autonomous actors to achieve common goals. If we consider open agent societies from an organizational point of view, we can identify a number of basic elements that design methodologies for agent societies should account for. These include formalisms for the description, construction and control of normative elements such as roles, norms and social goals, and mechanisms to formalize the expected outcomes of roles and to describe interaction in order to verify the overall behaviour of the society.

Interaction is one of the main elements of multi-agent organizational models, since it is the main—if not the only—way for agents to coordinate with one another (see the Chapter II “Modelling Dimensions for Agent Organizations” by Coutinho et al. for more information). In the literature, multi-agent communication has been the subject of a vast research activity addressing semantic and engineering aspects of multi-agent organizations. Two main approaches have emerged. In the early days of multi-agent research, a seemingly promising way to model agent interaction was largely inspired by Grice’s and Searle’s speech acts theory. This is now called the mentalistic approach since its focus is on the minds of the individuals participating in the interaction. Agent mental states would give motivation to message exchange, which in turn would affect the mental states of those participating in the exchange. This idea has been behind prominent Agent Communication Language (ACL) proposals such as KQML and FIPA-ACL. However, it became apparent that a semantics of agent communication based on mental states would necessarily impose significant restrictions on the architecture and operational behaviour of interacting parties, while making it difficult, at the same time, for an external observer to understand to what extent a message exchange would conform to such a semantics (Singh 1998).

Social approaches to agent communication seek to overcome these shortcomings and quickly gained large popularity. The idea of social semantics is to abstract away from the agent internals and provide a social meaning to agent message exchange. In other words, interaction is not motivated by the effect it may have on the mind of the agent, but instead on its visible effects on the agent’s social environment. Besides paving the way to the development of a number of techniques aimed to make agent interaction verifiable by external observers, social semantics have proven to be a viable approach to accommodate truly open agent societies, since they do not pose restrictions of any sort on the nature and architecture of interacting parties. These are key factors that made social semantics much more widely adopted than mentalistic approaches, especially in the context of organizational models.

The second aspect we mentioned relates to multi-agent systems engineering and design. Again, social semantics of agent interaction has been successfully applied to Multi-Agent System (MAS) design, both with respect to methodologies and formal reasoning about models. Current Agent-Oriented Software Engineering methodologies—see for example Gaia (Zambonelli et al. 2003), MaSE (DeLoach et al. 2001), Prometheus (Padgham & Winikoff, 2004) and the Hermes methodology presented earlier on in this book (see Chapter V “Hermes: A Pragmatic Approach to Flexible and Robust Agent Interaction Design and Implementation,” by Cheong and Winikoff)—include, at some design stage, modelling of actions, events, roles, normative relations and interaction protocols. It is possible to give a social semantics to these elements of a MAS by modelling them or their effect in terms of socially meaningful concepts. Modelling greatly benefits from the social semantics being declarative. A declarative, as opposed to procedural, semantics specifies what actions should be brought out in an interaction, rather than how they are brought out, and by doing so it helps to focus on the aims of the interactions and thus avoid designing unnecessarily over-constrained interaction patterns and modalities, as pointed out by Yolum & Singh (2002b).

Social semantics can be approached in two ways: by taking a state-oriented perspective, or a rule-oriented perspective. The main elements of these two approaches are, respectively, commitments and expectations. They both provide a formal, conceptually well-founded basis for modelling the interaction of agents while respecting their autonomy.