The interaction system enables the agent to perceive and acquire information from the surrounding environment. This individual perception contributes to the elaboration of a subjective, limited, and contextual representation of the world.

From a philosophical perspective, there are two conflicting theories on perception. On one hand, the Aristotelian view assumes that perceived objects actively ‘impregnate’ our senses. Thus, we passively receive this imprint and integrate it to our cognitive system. This causal conception of perception asserts that we can access the objective qualities of surrounding objects. This model is widely accepted in cognitive science (logic theory) and computer science (shape recognition). On the other hand, the Kantian view asserts that percepts are constructed by the observer and depend upon previous experiences of perception. This active conception of perception constitutes an axiomatic principle in semiotics and it is consistent with major experimental results in neurobiology. Unfortunately, its application to computer science raises several technical problems that have, so far, limited its use in AI despite valuable experiments such as the ‘Talking Heads’ (Kaplan 2001). Figure 3.2 presents two computational systems of perception for artificial agents based on active or passive perception.

Figure 3.2. Passive and active perception systems. (Adapted from Ferber 1999)

The system of representation enables the cognitive agent to store and manipulate acquired knowledge and beliefs. AI tends to group knowledge, know-how, experience, facts, and memories into a single set of information called ‘beliefs’. These beliefs help the agent to decide and to implement actions. As a matter of fact, much theoretical work has been concerned with ways of representing, classifying and manipulating these beliefs for action (pragmatic dimension) rather than focusing on the very essence of ‘knowledge’ (epistemological dimension).

The Physical Symbol System Hypothesis, enunciated by Simon and Newell, is the founding principle of Symbolic Artificial Intelligence (SAI). Borrowing concepts from philosophy of mind (Kantian schemata) and semiotics (Peircean symbols), the principle states that any belief can be represented through a set of symbols and rules of inference (ontology). Four axiomatic propositions are generally accepted:

Evidence coming from neuro-biology has supported criticism of the first proposition by Connection Artificial Intelligence (CAI). The use of neural networks for task-oriented reasoning has indeed provided powerful alternate solutions. But, a more general criticism towards SAI relates to its implicit assumption of perfectly autonomous agents. As a matter of fact, social agents are embedded into an environment that not only supports and feeds individual reasoning but, more essentially, ‘permeates’ individual experiences through permanent interactions. Social psychology asserts that intelligence is culturally grounded and that knowledge evolves only through interaction with others by means of proposition, confrontation, and refutation (Cole and Scribner 1974). Interestingly, this same criticism gives some credit to the unique concept of belief used by SAI to describe different types of knowledge: from a social psychology viewpoint, any type of knowledge results from a historically contingent and socially built consensus. Hence, scientific knowledge and theories are themselves meta-beliefs, consensual models of a given ‘reality’.

From an SAI perspective, agents continually use beliefs and assumptions (cognitons) from their system of representation to build descriptive or predictive models of their environment. Ferber (1999) proposes the following list of belief categories:

The conative system defines the set of activities to be undertaken by an agent, based on available information and beliefs. The ways agents take their decisions, and the reason why they discard some options to focus on others, are questions that stretch well beyond Artificial Intelligence and nurture endless debates in philosophy and psychology.

As SAI relies upon logic inferences to describe an agent’s behaviour, such as predicate logic or modal logic, causality links are meant to be rational. Hence, agents tend to display goal-satisfying decisions and, therefore, their actions are first driven by their needs and tendencies. According to Ferber (1999), these tendencies are themselves motivated by percepts, impulses, norms, or engagements, and trigger a decisional process based on existing beliefs and assumptions (Figure 3.3). Like beliefs, motivations can be separated into four categories:

Figure 3.3. Conative system and its two sub-systems. (Adapted from: Ferber, 1999)

The way SAI agents take intentional decisions and eventually undertake subsequent actions is largely based on causal philosophy of action (Bratman 1987). Intention is altogether a choice and an engagement towards this choice. Hence, in AI, an agent X is said to have the intention to perform an action A if X wants a proposal P about the state of the world to be true, and:

Cohen and Levesque (1990) have proposed a formalism for rational action, based on modal logic that has been largely used in Distributed Artificial Intelligence (DAI). Their formalism gives way to necessary, possible, or contingent predicates. Likewise, it allows expressing the temporality of intentions as planning to do something in the future, and needs a different set of tasks compared with deciding to do something now. Applying such formalism to Multi-Agent Systems implies that each agent is able not only to predict the consequences of its intended action, but also to anticipate the results of the other agent’s behaviour. Hence, the agent needs to carry in his social or relational beliefs some ideas about the other agent’s commitments. This is where the concept of engagement becomes paramount: self and social engagement are needed to introduce some sort of regularities in the system that can be hopefully anticipated.

Finally, the organisation system, through its methods and tasks (cognitons), allows the agent to prioritise, halt, and resume the pending decisions provided by the conative system. External information channeled through the interaction system may alter the implementation of a decision into action. A suspended decision will eventually resume according to the persistence of its triggering intention(s) (Ferber 1999).

It is one thing to understand how SAI is used to design rational and intentional cognitive agents. But we also have to question the reasons why, in the first place, we intend to create these artificial entities? I will leave aside DAI applications belonging to robotics or computer-oriented technologies where autonomous agents tend to ‘mimic’ intentional cognition in order to perform actions considered as rational by their designers. After all, these agents are not supposed to be, or even to represent, human beings.

Instead, I will concentrate on these Multi-Agent-Based Simulations (MABS) that are meant to represent actual human ecosystems. Only a small proportion of those applications are used as social virtual experiments to explore cognitive processes. Relying on robust cognitive architectures inherited from SAI, these models are designed to help theoretical breakthroughs:

I believe that the contribution of [Multi-Agent-Based Social Simulation] to the theoretical development of the cognitive and social sciences could be really remarkable. SS can provide not only an experimental method, but good operational models of cognitive ‘actors’, of the individual social mind, of group activity, etc. Models that can be richer, more various, and more adequate than those provided by economics, without being less formal. In particular, my focus on the core relation between functions and cognition was aimed at pointing out how the coming ‘agent-based’ approaches to social theory, using learning but deliberative agents, could deal with very old and hard problems of the social sciences and could re-orient them.
(Castelfranchi 2001: 35)

As a matter of fact, a large proportion of MABS applications are designed to explore and understand complex interactions between actual actors and their environment. Bousquet and LePage (2004) or Hare and Deadman (2004) provide comprehensive reviews of these models. Most of these applications depart from the SAI paradigm and implement over-simplistic, task-oriented, rule-based agents, focusing on spatial interactions, social communication and individual mobility. Often, the drift from internally consistent and Formal Logic Compliant (FLC) agents is justified by synthetic information coming from field surveys or expert knowledge.

I would argue that in both cases, formal logic compliance or not, MABS will fail to deliver if we cannot find innovative ways to link the model, its object, and its interpreter. For deceptive idols and socially constructed icons are conspiring against agents.