The use of the agent paradigm to understand and design complex systems occupies an important and growing role in different areas of social and natural sciences and technology. Application areas where the agent paradigm delivers appropriate solutions include online trading,16 disaster management,10 and policy making.11 However, the two main agent approaches, Multi-Agent Systems (MAS) and Agent-Based Modeling (ABM) differ considerably in methodology, applications, and aims. MAS focus on solving specific complex problems using autonomous heterogeneous agents, while ABM is used to capture the dynamics of a (social or technical) system for analytical purposes. ABM is a form of computational modeling whereby a population of individual agents is given simple rules to govern their behavior such that global properties of the whole can be analyzed.9 The terminology of ABM tends to be used more often in the social sciences, whereas MAS is more used in engineering and technology. Although there is considerable overlap between the two approaches, historically the differences between ABM and MAS are often more salient than their similarities. For example, it is often remarked that a main difference between ABM and MAS is that ABM models are descriptive aiming at explanatory insight into the collective behavior of agents at the macro level, whereas MAS are operational systems, acting and affecting its (physical) environment, with a focus on solving specific practical or engineering problems, and emphasizing agent architectures with sophisticated reasoning and decision processes. This has lead to the development of two research communities proceeding on nearly independent tracks.
However, this division is not as black and white as it may seem. In fact, much ABM work goes beyond descriptive simulations of a situation, and, as input for decision making and policy setting, indirectly affects the environment. And, the design of MAS is often geared to analytic insights and simulations toward the understanding of how configurations of agents behave in different circumstances. Currently, applications of MAS are broader than pure distribution problems, including interactive virtual characters, where the focus is on the cognitive, affective, and emotional characteristics of the system, and game-theoretic models, focusing on the design of incentive mechanisms that guarantee a given strategic behavior.
Social abilities are central both in ABM, where agents represent humans and their interactions, and in MAS, that enable game-theoretic analyses of decision strategies, or provide interactive virtual agents in varied situations. It is precisely in this area where the need for integration of ABM and MAS is undoubtedly the most necessary. In social simulation, the benefits of combining MAS and ABM have been advocated for many years, and are the focus of the long-lasting workshop series on Multi-Agent Based Simulation (MABS).2 ABM has increasingly and successfully been used for social simulations,3 but it is in the MAS area that fundamental research on agent architectures implementing psychological traits and social concepts such as norms, commitments, emotions, identity, and social order, has been most prominent.4,5 Bridging these somewhat parallel tracks requires a new grounding for agent architectures.
Questioning Rationality
Traditionally, one of the most salient aspects shared by both ABM and MAS approaches is the premise of rationality. This is derived from the traditional definition of agents as autonomous, proactive, and interactive entities where each agent has bounded (incomplete) resources to solve a given problem; there is no global system control; data is decentralized; and computation is asynchronous.21 Agent rationality can be summarized as follows:
- Agents hold consistent beliefs;
- Agents have preferences, or priorities, on outcomes of actions; and
- Agents optimize actions based on those preferences and beliefs.
This view on rationally entails that agents are expected, and designed, to act rationally in the sense that they choose the best means available to achieve a given end, and maintain consistency between what is wanted and what is chosen.14 Even though multiple alternatives have been proposed, in both the ABM and MAS approaches, individual agents are still typically characterized as bounded rational, acting toward their own perceived interests. The main difference is that agent behaviors in ABM are used to capture the dynamics of a system for analytical purposes, grounded whenever possible on existing data about system outcomes, whereas MAS focuses on solving specific problems using independent agents, through the formalization of the complex goal-oriented processes, such as the Beliefs-Desires and Intentions (BDI) model proposed by Bratman20 or game-theoretic approaches.
Unfortunately, from a modeling perspective, real human behavior is neither simple nor rational.
The main advantages of such rationality assumptions are their parsimony and applicability to a very broad range of situations and environments, and their ability to generate falsifiable, and sometimes empirically confirmed, hypotheses about actions in these environments. This gives conventional rational choice approaches a combination of generality and predictive power not found in other approaches. In fact, rationality approaches are the basis of most theoretical models in the social sciences, including economics, political science, or social choice theories.
Unfortunately, from a modeling perspective, real human behavior is neither simple nor rational, but derives from a complex mix of mental, physical, emotional, and social aspects. Realistic applications must consider situations in which not all alternatives, consequences, and event probabilities can be foreseen. This renders rational choice approaches unable to accurately model and predict a wide range of human behaviors.
Toward Social Agents
Human sociability refers to the nature, quantity, and quality of interactions with others, including both pro-social, or cooperative, behaviors, and conflict, competitive, or dominating behaviors. Sociability is also the ability to influence others, by changing their behaviors, goals, and beliefs, the emotional reaction to others and to the environment, and how actions are affected by emotions, and the ability to create, structure and ‘rationalize’ the environment to fit ones expectations and abilities (leading, for example, to the design of organizations, institutions, and norms).
Following an increasing number of researchers in both ABM and MAS that in recent years have come to similar conclusions,7,13,18,19 we claim that new models of preference and belief formation are needed that show how behavior derives from identities, emotions, motivation, values, and practices.6
The endeavor required to construct such agent models that are socially realistic requires the effort and the capabilities of both the MAS and ABM communities, bringing together formalization and computational efficiency, and planning techniques as in MAS, with the ABM expertise on empirical validation and on adapting and integrating social sciences theories into a unified set of assumptions,1 furthering the fundamental understanding of social deliberation processes, and developing techniques to make these accessible for simulation platforms. This Viewpoint is therefore an appeal to join the strengths of both communities toward sociality-based agents.
Without claiming a readily available solution, we propose the concept of sociality as the leading principle of agency, as an alternative for rationality. Following the aforementioned description of rational behavior, the main characteristics of sociality-based reasoning are:
- Ability to hold and deal with inconsistent beliefs for the sake of coherence with identity and cultural background. That is, beliefs originate from other sources than observation, including ideology or culture.
- Ability to fulfill several roles, and pursue seemingly incompatible goals concurrently, for example, simultaneously aiming for comfort and environmental friendliness, or for riches and philanthropy.
- Preferences are not only a cause for action but also a result of action. Moreover, preferences change significantly over time and their ordering is influenced by the different roles being fulfilled simultaneously, which requires the need to deal with misalignment and incompatible orderings.
- Action decisions are not only geared to the optimization of own wealth, but often motivated by altruism, fairness, justice, or by an attempt to prevent regret at a later stage.
- Understand when there is no need to further maximize utility beyond some reasonably achievable threshold.
- Understand how identity, culture, and values influence action, and use this knowledge to decide about reputation and trust about who and how to interact.
The first step toward sociality-based agents is a thorough understanding of these principles, and open discussion across disciplines on the grounds and requirements for sociality from different perspectives. This discussion will be fundamental to the development of formal models and agent architectures that make sociality-based behavior possible and verifiable.
Moreover, it is necessary to identify and formalize which mechanisms, other than imitation, can describe how agents can adapt to pressures in the environment to behave in a socially acceptable, resource-sustainable fashion. Resulting models support the understanding or predicting human behavior, including rich models of emotions, identities, culture, values, norms, and many other socio-cognitive characteristics. Such models of social reality are also needed to study the complex influences on behavior of different socio-cognitive characteristics and their relationships. The integration of psychological models of motivation and cognition, sociological theories of value and identity formation, and philosophical theories of coherence and higher-order rationality, together with different formal methods, quickly yields intractable models. However, it is important to identify what is the model being developed for. In fact, richer models are not always the most appropriate ones.
Sociality-based agents are also fundamental to the new generations of intelligent devices.
Once these characteristics are well understood, then simplified models can be developed to suit different needs. That is, implementing sociality-based agents will require other techniques than those currently used in either MAS or ABM,8 including the use of simpler, context-specific decision rules, mimicking how people themselves are able to deal with complex decision making, for example, using social practices as a kind of shortcuts for deliberation.15,17 Where it concerns utility, satisficing can be more suitable approach than maximizing.12 This also allows us to integrate agents of varied richness levels, for example, using rich cognitive models to zoom-in the behavior of salient agents in a simulation, whereas other agents just follow simple rules. This approach can counter the obvious criticism that sociality-based agents will become too complex for use in computational simulations.
Sociality-based agents are also fundamental to the new generations of intelligent devices, and interactive characters in smart environments. These artifacts not only must build (partial) social models about the humans they interact with, but also need to take social roles in a mixed human/digital reality. An interesting challenge would be to use the same technologies in real time mixed human/artificial interactions, and criticisms could also be on the feasibility to use these architectures (or controlled reductions/simplifications) in real time or near real time.
Moving Forward
The intent of this Viewpoint has been to appeal for a collaborative research effort toward fundamental formal theories and models that increase our understanding of the principles behind human deliberation (such as the ones listed discussed here), before deciding on which modeling techniques we need to implement them. Even though, several approaches to model social aspects in agent behavior are available, there is not sufficient consensus on which characteristics are needed for what, nor on how to specify and integrate them. We have identified an initial set of characteristics for sociability, proposed a research path linking theory, model, and implementation, and suggested possible theories and techniques to develop sociality-based agents. These incorporate expertise from both ABM and MAS and require integration of both areas in order to succeed. We welcome the discussion of these issues toward a novel area of research on social agents, which take sociability as the basis for agent deliberation and enable interaction.
Join the Discussion (0)
Become a Member or Sign In to Post a Comment