Beside the symbolico-cognitive paradigm, Multi Agent Systems explicitly emphasise the autonomy of agents through physical integrity, autonomous decision, and rational action. But social psychology (Cole and Scribner 1974), psychology of development (Papert 1980) and collective action theory (Oliver 1993) insist, for different reasons, on the fact that human beings are social beings before anything else: learning processes, cognitive inferences, or individual actions are shaped and dictated by our environment. Recent developments in DAI have tried to incorporate this ‘sociality’ dimension into agent’s behaviour. We’ll come back in the next section to Hogg’s and Jennings’ (2001) proposal to include social rationality into agents’ expected utilities, or Jager’s and Janssen’s (2003) attempt to consider basic human needs and uncertainty as the driving factors behind decision making processes of their agents. But first, we need to understand the elements of social cognition that weaken the foundations of the SAI paradigm.
Charles Sanders Peirce, a founder of modern semiotics, has asserted that:
we cannot think without signs,
we have no ability for intuition, all knowledge flows from the former knowledge,
we have no ability for introspection; all knowledge about the inner world is produced by hypothetical reasoning on the basis of observation of outer things. (Peirce Edition Project 1998)
On this solid foundation he builds up his entire theory of signs. Being a pure positivist himself, at least during his early career, Peirce is convinced to build a theory based on formal logic, independent from particular minds, and characterised by a triadic relation (Peirce Edition Project 1998: 411):
[semiosis is an] action, or influence, which is, or involves, a cooperation of three subjects, such as a sign, its object, and its interpretant, this tri-relative influence not being in any way resolvable into actions between pairs.
Peirce suggests that people have not and cannot have direct access to reality. Signs are nothing else than the universal medium between human minds and the world. Semiotics defines different categories of signs according to their level of abstraction, from icons to symbols (Peirce Edition Project 1998). As a primary and elementary sign, an icon looks like it is signified. There is no real connection between an object and an icon of it other than the likeness, so the mind is required to see the similarity and associate the two. A characteristic of the icon is that by observing it, we can derive information about its significance. Hence, semiotics embrace the Kantian concept of active perception as stated previously.
Since signs are not private but socially shared, it is society that establishes their meaning. Therefore, the transcendental principle in semiotics is not (divine) intuition, but community, and the criterion of truth, social consensus. According to Peirce, any truth is provisional, and the truth of any proposition cannot be certain but only probable. Umberto Eco, in his exhilarating Kant and the Platypus, provides a vivid description of this fallibilism principle (Eco 1997: 98):
Naturally at this point transcendentalism will also undergo its Copernican revolution. The guaranty that our hypotheses are ‘right’ will no longer be sought for in the a priori of the pure intellect but in the historic, progressive, and temporal consensus of the Community. Faced with the risk of fallibilism, the transcendental is also historicized; it becomes an accumulation of interpretations that are accepted after a process of discussion, selection, and repudiation. This foundation is unstable, based on the pseudo-transcendental of the Community.
In order to sort out conceptual confusions born from consensual uncertainties, Pierce uses a ‘pragmatic’ approach by linking the meaning of concepts to their operational or practical consequences. This pragmatism will act later on as a corner stone for SAI’s intentionality of beliefs. But, when conventional SAI says: ‘beliefs are intentional’, semiotics tell us: ‘we have to consider beliefs as intentional’. Somehow, SAI has developed an hyper-positivist approach of cognition, through passive perception, intentionality, and formal modal logic that has gone beyond the scope of its theoretical background. In this regard, FLN agents are more likely to ‘mimic’ social behaviour compared with their FLC counterparts as the rationale for their actions is inferred from an holistic, though subjective, view of the social system constructed by their designer. In other words, some behavioural rules have to be considered for the FLN agents to interact in a consistent way (system level), without prerequisite conditions on cognitive consistency (agent level). But we’ll see in the next section that it comes at a cost.
We human beings are living systems that exhibit cognition, and there is no way for us to address, much less explain, our cognitive abilities without employing those same cognitive abilities. To date, the primary response to this paradox has been to ignore it and proceed with respect to a presumably fixed fundament, external to our act(s) of cognition. Where the presumptive fundament is an ‘objective reality’, the mediation between situation and action is explained in terms of ordered inference with respect to a model of that reality.
Maturana and Varela (1980) questioned this conventional approach when it is confronted with the tacit, extralinguistic, or emotive character of human behaviour. They further disputed the ‘objective reality’ concept by considering:
Ourselves operating in multiple ‘worlds’, particularly socio-cultural ones.
A ‘world’ being molded by contextual factors intertwined with the very act of engaging it.
Their autopoietic theory considers living beings as living systems embedded into larger systems constituted by themselves and the environment they interact with. The theory addresses the basic configural and operational circularities of these living systems (Maturana and Varela 1980). Unlike other more positivist approaches to complex systems, the autopoietic theory focuses on the observer himself. It accomplishes this by shifting explanatory focus from atomic units in an objective world to essential relations among processes operating in circular ways to constitute the organism as a living system and the observer as a cognitive organism. As clearly stated by Maturana (Maturana and Varela 1980: 7):
Cognition is a biological phenomenon and can only be understood as such; any epistemological insight into the domain of knowledge requires this understanding.
Later on, Varela and colleagues brought phenomenological concerns into the world of cognitive science (Varela et al. 1991). Their goal was to incorporate everyday experience into the scope of studies which had heretofore addressed cognition in terms of disembodied rational processes, circumscribed by abstract beliefs purported to mirror an objective milieu. All concepts fully accepted by symbolico-cognitivist approaches. Varela and colleagues proceed from the assumption that experience necessarily predates and underpins enquiry. Maintaining a focus on experience as action allows inspection and reflection on the manner in which ‘mind’ and ‘body’ reciprocally engage to consummate experience. Midway between cognitivist and connectionist paradigms, their Enactive Cognitive theory considers that (Varela et al. 1991: 148):
context-dependent know-how [shouldn’t be treated] as a residual artifact that can be progressively eliminated by the discovery of more sophisticated rules but as, in fact, the very essence of creative cognition…Knowledge depends on being in a world that is inseparable from our bodies, our language, and our social history—in short, from our embodiment.
As such, enactive cognitive theory does not address cognition in the currently conventional sense as an internal manipulation of extrinsic information or signals, as semiotics or symbolico-cognitivism would have us believe (we have seen above how Maturana and Varela treat linguistic interactions) . Instead, it grounds cognitive activity in the embodiment of the actor and the specific context of activity. The theory fits very well with current trends toward emphasising contextualised studies of humans, their interactions, and their social systems. Unlike symbolico-cognitivist approaches, the enactive cognitive theory is consistently ‘relativistic’ in the sense that any given observation is observer-bounded, history-contingent, and socially embedded.
The theory implies an epistemology analogous to that of constructivism (Funtowicz and Ravetz 1993). For this reason, Multi Agent Systems built from FLN agents and based on expert knowledge never describe the reality as it is but rather a probable reality interpreted by a given observer in a given context. It means that such models cannot be evaluated through traditional scientific methods but need to be assessed, in context, against a set of criteria intelligible to the observed subjects and to the observer.
As mentioned above, very influential studies in social sciences have focused on the embodiment of cognitive activities into the actor’s experiences. In the psychology of development, Piaget asserted that children only develop through progressive interactions with their environment (Papert 1980). In the early stages of development, Piaget’s theory suggests that learning comes only from unexpected changes in the environment of the subject. Changes create a cognitive imbalance that must be adjusted through assimilation, the new experience is integrated into the current view of the world, or accommodation, the current view of the world must be modified to fit in the new experience. From this constructivist viewpoint, imbalance is no longer considered as a negative factor but rather as a driving force towards cognitive development. Hence, a perceived environmental unstability seems to trigger some sort of phase transition, in a complexity theory sense, from one state of representation to another. It is interesting here to draw a link with the enactive cognitive theory, for which concrete experiences (know-how) are driving cognitive processes through connotative interactions (Maturana and Varela 1980).
Indeed, a new picture of cognitive processes takes shape; a mind submitted to perpetual changes and characterised by a dynamic equilibrium between past and current views of the world; a mind dominated by inductive inferences and desperately trying to make sense of incoming information through deductive and formal logic (Batten 2000). This description sharply contrasts with the well structured and perfectly organised mind proposed by SAI designers. As a matter of fact, FLC agents process information in a consistent way; their actions are driven by intentional cognition and the experienced outcomes are evaluated against expected utilities (Brazier et al. 2002).
Marvin Minsky, in his famous Society of Mind (1985), proposed an alternate solution to the symbolico-cognitivist theory. Drawing from an initial intuition that our mind is made of structural and independent units called frames, Minsky views the human mind itself as a vast society of individually simple processing agents. The agents are the fundamental thinking entities from which minds are built, and together produce the many abilities we attribute to minds. The advantage in viewing a mind as a society of agents, as opposed to the consequence of some formal logic-based system, is that different mental agents can be based on different types of processes with different purposes, ways of representing knowledge, and methods for producing results. The consistency of this atomistic mental system rises from the organisation of interactions between agents, not from a unified and coherent reasoning (Minsky 1985: 308):
What magical trick makes us intelligent? The trick is that there is no trick. The power of intelligence stems from our vast diversity, not from any single, perfect principle.
In principle, these networks of mental agents display dynamic interactions where agents can be modified, created, or deleted in order to adapt or to respond to a given experience. This interactionist theory can accommodate—though not explicitly formulated—constructivist (Papert 1980) and connotative (Maturana and Varela 1980) standpoints. Unfortunately, to date, few Multi Agent Systems have been implemented according to Minsky’s views, and a vast majority of these models concern robotics or computer-oriented technologies. Both fields have been excluded from the scope of this paper.
Another advantage of the interactionist theory is that it creates a continuum between individual and social actions. The same paradigm is used from mental processes to collective action, from personal engagement to societal norm enforcement. As a matter of fact, sociologists studying social movements often refer to Collective Action Frames to represent the process by which activists create and modify action-oriented sets of beliefs and meanings that inspire and legitimate the activities of a social movement (Benford and Snow 2000). A crucial aspect of collective action frames is that they are not merely aggregations of individual attitudes and perceptions but also the outcome of negotiating shared meaning. Collective framing necessitates complex interactions between activists, antagonists, bystanders, and observers (op. cit.: 614):
Frames help to render events or occurrences meaningful and thereby function to organize experience and guide action. Collective action frames also perform this interpretive function by simplifying and condensing aspects of the ‘world out there’, but in ways that are intended to mobilize potential adherents and constituents, to garner bystander support, and to demobilize antagonists.
Cederman’s quest for generative explanations of social forms based on configurative ontologies cannot be entirely satisfied by autonomous and rational agents, as formal modal logic constitutes a very specific and limited type of ontology. As we have seen, the interactionist paradigm offers a much richer context for cognition and sociation. Unfortunately, technical caveats have prevented so far to implement purely interactionist computer models beyond experimental and limited applications. Finally, Cederman’s call to the Simmelian tradition questions the nature of these social patterns we are supposed to observe: if we accept Maturana’s and Varela’s views, these patterns are strongly observer-dependent and they need a constructivist epistemology for authentication.