Rosaria Conte and Frank Dignum (2001)

From Social Monitoring to Normative Influence

Journal of Artificial Societies and Social Simulation vol. 4, no. 2,
<https://www.jasss.org/4/2/7.html>

To cite articles published in the Journal of Artificial Societies and Social Simulation, please reference the above information and include paragraph numbers if necessary

Received: 01-Nov-00      Accepted: 01-Feb-01      Published: 31-Mar-01

Abstract

This paper is intended to analyse the concepts involved in the phenomena of social monitoring and norm-based social influence for systems of normative agents. These are here defined as deliberative agents, representing norms and deciding upon them. Normative agents can use the norms to evaluate others' behaviours and, possibly, convince them to comply with norms. Normative agents contribute to the social dynamics of norms, and more specifically, of norm-based social control and influence. In fact, normative intelligence allows agents to

1. Check the efficacy of the norms (the extent to which a norm is applied in the system in which it is in force), and possibly
2. Urge their fellows to obey the norms.

The following issues are addressed:

• What is norm-based control?
• Why and how do agents exercise control on one another?
• What role does it play in the spread of norms?

Keywords:

norms, multi agent systems, imitation, social control, social cognition

Introduction

1.1

As is illustrated by some authors (Tambe 1997; Kaminka & Tambe, 1998), coordination between agents^[1] in a multi-agent system is far from trivial. Although it is possible to have agents work on a collective plan or cooperate (and even compete) according to some established protocol, strange things happen when one of the agents does not fulfil her part of the bargain (maybe due to some external circumstances). For example, a helicopter disrupts the complete plan by flying into enemy territory by accident, because it misses a delimitation point. The other helicopters do not notice this because their part of the plan is not (directly) disrupted and they are not monitoring (controlling or checking) the behaviour of the other helicopters. This shows that a cooperative plan might benefit greatly from the agents' capacity to check one another, thereby preventing misdoings, which might compromise the collective result even without impairing the others' activity.

1.2

It seems that an agent that is part of a multi-agent system should also be aware of the other agents in the system and of the relations that exist between the agents and the rules that govern these relations. These rules and relations determine the balance between the autonomy of the agent and the inter-dependencies between the agents. In human societies, these rules are determined in large part through (formal or informal) social norms. They form a very subtle mechanism according to which humans can be members of social groups, but at the same time autonomous. They can compete with other humans (as in commerce) while at the same time obeying some social rules (like not committing fraud). The main thesis of this paper is that the normative social system that exists in human society should function as a starting point for constructing social rules in multi-agent systems. Only when we have some understanding about the human social system can we determine which parts are relevant for multi-agent systems and how they should be implemented.

1.3

This paper is intended to give a first analysis of the concepts involved in the phenomena of social monitoring and norm-based social influence for multi-agent systems with normative agents. Normative agents are here defined as Belief/Desire/Intention agents (that is, agents endowed with a set of mental properties necessary for deliberating) deciding upon norms. Throughout this paper, these properties will be referred to as normative intelligence. Although the goal is to use the concepts found in multi-agent systems, the analysis is mainly based on human societies and studies in the social sciences.

1.4

This work is part of a research project (cf. Conte et al., 1998) on the study of the emergence and spread of norms in intelligent agent societies, that aims to implement norms in BDI agents and run simulation experiments to check the global social advantages of normative intelligence. Such advantages may include an intelligent application of norms, the solution of conflicts between norms, the execution of contrary-to-duty obligations. A special effect of normative intelligence at the global level is norm-based social control. Normative intelligence enables agents to monitor one another's behaviours against existing norms. Agents endowed with normative intelligence (from now on, normative agents^[2]) can use the norms they have accepted to evaluate others' behaviours and, possibly, convince their fellows to comply with norms. More generally, we have argued (Conte et al., 1998) that a theory of normative intelligence may contribute in two important ways to a scientific study of norms.

1.5

First, it may bridge the gap between legal and social norms. In AI, there are two main fields where normative concepts and phenomena abound:

• Logic philosophy (deontic logic, social philosophy, philosophy of law) and its computational application, especially in the areas of legal expert systems, normative reasoning and diagnosis;
• Social sciences and agent systems (multi-agents systems, cooperative systems, autonomous agents modelling) and related computational applications, especially in the areas of Computer Supported Cooperative Work and electronic commerce.

1.6

A wide gap exists between these fields. They differ in many ways^[3]. One important difference lies in their respective concept of a norm, which is mainly interpreted as a legal norm in the former field, and as a social convention in the latter field. In our view, such a gap may be filled thanks to a general theory of norms, whether legal or social, as prescriptions impinging on agents' minds before and in order to regulate their actions. Whether a norm is institutionally issued or emerges from interactional practice, it will actually rule the action of intelligent deliberative agents if they are enabled to recognise it and decide to obey or reject it. Norm-based decision-making is essentially common to both legal and social norms, while of course they differ in their origins and dynamics. A fundamental step common to both types of decision-making is that of norm-recognition, which was regarded by some philosophers of law (Hart 1961) as crucial for legal norms. Agents that recognise norms tend to enforce and propagate them in their social environment. More or less deliberately, they act as norm issuers. Both legal and social norms profit from this non-institutional propagation. Thanks to normative intelligence, the difference between the social dynamics of legal and social norms is fairly attenuated. Another important aspect of the relation between these two sub-sets of norms, legal and social, is to what extent they may contribute to each other's development or evolution. Can social norms lead to the formation of legal norms, and can legal norms lead to social norms being established, and if so, over what time-scale? This paper is not the forum for addressing such a fundamental issue (but see again, Conte & Castelfranchi 1999, for a conceptual discussion of norms intended to bridge the gap between legal and social norms). However, it should not be forgotten that some crucial meta-norms (like the norm of reciprocity) are plausibly at the origin of most social and legal norms (see Ostrom 1998). This is not to say that legal and social norms necessarily share the same roots, since many legal norms may be issued for rather arbitrary and technical reasons. However, some normative principles, like reciprocity, seem to operate in both domains. Furthermore, legal norms may, and often do, generate new social values. The new legislative regulation of nicotine and smoking behaviour introduced in western countries has undoubtedly not only favoured a more responsible behaviour on the part of smokers with regard to their social surroundings, but is also associated with a different attitude with regard to people's own health and individual well being, which might be perceived as a new socially shared value evolving. Whether this is at the origin of the new legislation or a consequence of it is hard to say, although a co-evolutionary and interactive process seems more plausible.

1.7

Secondly, a theory of normative intelligence may help to understand several aspects of the social dynamics of norms, and more specifically:

• Norm evolution: agents' capacity to reason about, recognise and decide whether to accept or not, to observe or not, a norm may give insight about the process through which a norm is selected and possibly is modified. Autonomous and intelligent norm-based decision-making contributes to the evolution of norms no less than it contributes to the evolution of other social constructs. If norms are implemented as mere action constraints in multi-agent systems, they cannot evolve since agents are not allowed to violate them. Autonomous action allows violation, which in turn allows norm evolution.
• Norm formation. Analogously, normative intelligence is involved in the process that gives rise to a new norm. Normative agents may not only decide whether to observe or violate existing norms, thereby reinforcing or extinguishing them, but also decide whether a world state is or should be a norm. The latter decision, which we have called autonomous norm formation (see again, Conte et al. 1998) , involves a set of mental operations upon normative representations and other mental states, including instantiating existing norms and generating new norms from pre-existing meta-norms or principles.
• Norm-based social control and influence. Normative intelligence allows agents to
  1. Check the efficacy of the norms (the extent to which a norm is applied in the system in which it is in force) by monitoring one another's behaviours against the norms they have accepted, and possibly to
  2. Urge their fellows to obey the norms.

1.8

In this paper, we will focus on normative control and influence, while the contribution of normative intelligence to norm formation and evolution will be discussed in future work. More specifically, we will address the following issues:

• What is norm-based control? Are there different ways in which it is executed?
• Why do agents exercise this type of control on one another, and which are the mental processes involved?
• What role does it play in the spread of norms?

1.9

This paper is only a first step to work out a complete theory, and is mainly aimed at making explicit hypotheses for further computational and simulation work. Results will consist of possible contributions to:

• Agent and multi-agent theory and architecture. Promising applications of agent systems abound in both cooperative (e.g., personal assistance and teamwork) and competitive (e.g., electronic commerce) domains. Both show an increasing demand for a capacity to form or instantiate informal social precepts (such as reciprocity or respect of privacy) and other "collective" solutions (for some examples, see the role of reputation in Chavez et al. 1997). Apparently, these applications would greatly profit from a theory of how normative intelligence can contribute to the dynamics and evolution of norms.
• Social scientific theory of norms and social dynamics: Computational methods for the study of social dynamics, cultural evolution, and "memetics" are on the increase. Social norms and conventions have long represented a phenomenon of interest for computational social scientists and a test-bed for the application of their methodology. On the other hand, agent-based simulation of social phenomena is a rather promising sub-field. Integrating norms into the agent models for social simulation would enable these models to account for the difference between the regularities which propagate because they are socially prescribed and those which propagate independently or despite existing social prescriptions.
• Modelling social complexity. A theory of normative intelligence may provide fundamental insights for the study of complex social systems, and especially for modelling the co-evolution of mental and social complexity. Social complexity is implemented on mental complexity, but at the same time mental complexity may reduce social complexity. The capacity to accept and violate the norms has a strong influence on both the regulation of social systems and on the evolution of social regulation.
• Social-economic development and policy-making. It is impossible to introduce new policies, and design and enforce institutions aimed at governing existing social-economic processes without investigating what type of effect these will generate on the agents' minds, and more crucially, which mental properties and phenomena could cooperate with the candidate policies and institutions. A general theory of normative intelligence paves the way for such an investigation.

1.10

In the next section, norm-based control will be investigated as a multi-level, evolutionary notion. It will be shown to be rooted in social monitoring and possibly lead to normative influence. In the following section, the concept of social monitoring will be examined in its main components: motivations, mental ingredients and effects. In section 4, social control will be defined as a step in the process of social influence, and its motivations, requirements and effects analysed. Interestingly, social control may not lead to the exercise of any influence, but to reconsidering one's own norms. The interesting question about when social control leads to social influence will be raised in section 5, where some answers will be discussed, and the mental ingredients and effects of normative influence will be analysed. In the final section, directions for future works will be pointed out.

From Social Monitoring to Social Influence

2.1

One of the most influential views derived from social psychology states that social phenomena propagate thanks to imitation of others' behaviours (Bandura 1977). This view has had a vast influence in several social scientific fields, especially in evolutionary game theory (Weibull 1996) and in social dilemmas theory (Liebrand and Messick 1996). Conventions, norms, regularities, social structures and patterns of any sort (segregation, discrimination, opinion formation) are said to emerge from agents' imitation of either the most successful, or the most frequent, or simply the closest behaviours.

2.2

In the multi-agent systems field, several attempts have been made to check the distribution of conventions and social laws thanks to one or other of the above imitative strategies (Shoham and Tennenholtz 1995; 1997; Walker and Wooldridge 1995, etc.).

2.3

But imitation is not a sufficient explanation of social dynamics. Agents not only imitate passive models, whose influence is non-deliberate and even inadvertent, but also respond to active influence, deliberately exercised. Why do agents influence one another? So far, the main answer to this question (cf. Homans 1951; 1974) is in terms of a behavioural law of reinforcement: agents provide social approval or disapproval because they want to obtain approval in return or avoid disapproval. However, the reinforcement law cannot do justice to the complex mental processes that lead agents to influence one another. To see this, suffice it to consider that agents are able to discriminate between a normative and a non-normative influence, even if yielding to both will gain them some sort of approval, or any other type of reinforcement. Where does such a discerning capacity come from? Looking at the same phenomenon from the other side of the coin, agents are able to urge others to either conform to given behaviours or simply obey the norms, whatever they are. How could they assume such a discerning capacity in others, if they had none?

2.4

In the following, the concept of imitation is replaced by the concept of social monitoring, defined in its weakest and most general form as the agents' attitude to observe and use others' behaviours as a criterion for their own behaviours. Within this general process, we will pick one specific form of social monitoring, aimed at detecting existing norms.

2.5

Norm-based social control will be analysed as a subsequent step (Figure 1). Here, others' behaviours are monitored in order to check whether norms are effectively executed. This may be due to several reasons, only some of which, under given conditions, will lead to the further, conclusive step, namely norm-based influence. The reasons and ingredients of this process will be analysed in some detail at each of the steps of the process thus described.

Figure 1. From monitoring to influence

2.6

The general idea underlying our approach is that either agents use others' behaviour as a criterion, or they use norms as a criterion against which to evaluate others' behaviours, and in this case, they may try to modify it (norm-based influence) or not. In this paper, we will not go into the issue of norm formation. However, we would like to mention that the monitoring of other agents' behaviour against some average or standard behaviour might lead to the actual formation of a social norm that "abstracts" the average (or accepted) behaviour into a separate concept. Social control may lead one agent either to influence others to observe a norm or to its rejection. The question is when does norm-based social control lead to normative influence.

Social Monitoring: A Multi-faceted and Scalable Notion

3.1

We will speak of social monitoring when one single agent observes and registers the behaviours of other agents belonging to a given group in order to adjust its own mental attitudes with regard to the norms or values or other standards in force among those agents. More specifically, an agent x "monitors" a set of agents Y (whether x is a member of Y or not), when it compares its behaviour, beliefs, goals, values, to those most frequently shown by the members of Y in order to find out similarities or discrepancies. In this sense, x is using Y's behaviours as a criterion for evaluating its own. The issue of social comparison is well-known to social psychologists. According to Festinger (1954), (social) agents have a tendency to compare themselves with their social environment in order to check possible discrepancies which might produce also discrepancies among their own cognitions (for example, her own values and the values of significant others). This view of social comparison is essentially belief-driven, and is based upon the intuition that agents need internal consistency.

3.2

The view of social monitoring proposed in this paper is rather action-, or goal-driven, and is based on the pragmatic assumption that agents turn to others for inputs under condition of imperfect or uncertain knowledge or instructions about how to behave in given conditions. In this conception, there are two possible goals of social monitoring.

• To check a belief or expectancy about a standard behaviour or norm of the group Y. Are the members of group Y really (or still) adhering to the believed standard?
• To test a hypothesis about a possible new norm or standard for group Y, i.e. to form a new norm.

3.3

In the following sections we will analyse the following aspects of social monitoring:

• Why does social monitoring occur? When and how often do agents resort to social monitoring?
• How is it realised? What are the cognitive representations and processes involved?
• What are the consequences of this phenomenon both at the group level and at the agent level?

Motivations

3.4

What are the agent's motivations to monitor its social surroundings? One immediate answer is suggested by the social learning literature (Bandura 1977): agents want to imitate others, or at least conform to those that they consider as valuable social models.

3.5

The follow-up question then is why do agents want to imitate others, and conform to their behaviours? Again, the classical social-psychological answer (Homans 1951; 1974) says that agents imitate others to obtain approval from them (see Figure 2).

Figure 2. Social learning

3.6

However, we believe that this answer accounts only part of the phenomenon. In other words, we think that the picture should include at least another path (see Figure 3).

Figure 3. Norm-oriented social monitoring

3.7

For example, agents might learn from one another what the consequences of norm violation are, that is punishment. In such a case, social surroundings do not motivate the individual agent to comply with the norm. Rather, they furnish the agent with a sort of experimental " testbed "^[4] (see Conte & Paolucci 2001). Examples of this sort abound in social life: agents not only observe and learn given behaviours from one another, but also avoid the costs of a direct experiment, and learn the positive or negative (side-)effects of current plans/procedures. Agents find positive or negative models that allow them to instantiate their own standards, or they may infer standards and either accept or reject them. Obviously, to look for social approval can be considered as a precursor of the will to conform, and imitation as a forerunner of conformity. However, once norms and institutions have emerged, they allow for other cognitive objects to emerge, for example the expectation that norms exist and must be fulfilled. The fundamental difference between the models in Figures 2 and 3 is that norms in the first model are defined as the standard behaviour of a group. They have no existence of themselves. Therefore the only way for an individual to adjust to group behaviour is to adjust to the standard group behaviour. But what about all other norms, which may not coincide with the group standard behaviours? People know that safety belts should be fastened while driving even if such a norm is actually violated by a good proportion of the population if not the majority. The efficiency of norms would be greatly reduced if norms could be only deduced from standard behaviours.

3.8

Through social monitoring, agents find a fundamental source of information about which specific norms are impinging on a given set of agents. They will observe others to find out which are the norms in force in that particular social context. In case discrepancies between themselves and others are found, they may decide that others apply a set of norms that they were not aware of. Consider the case of a newcomer entering a social setting. How will she know what rules should she follow? One fundamental way to get this information without being perceived as an outsider is to observe others.

3.9

To sum up, we will speak about x social monitoring a given set of agents Y, when either

• x checks for discrepancies between itself and Y in order to adjust itself to Y (see our definitions of behavioural and goal conformity in Dignum & Conte 1997) - in such a case x has Y's behaviours, preferences, as a criterion for reducing discrepancies;
• x checks Y in order to find out instances of standards (possibly normative and moral ones) - in such a case, x has abstract standards (to be instantiated) as a criterion for reducing discrepancies.

3.10

We believe that only this second case gives rise to x' s decision to reduce the discrepancy by inducing members of the group Y to change their behaviour. Only when x has an independent norm to which she believes the members of Y ought to comply, does she have the motivation to persuade them to observe that norm.

Mental requirements

3.11

The difference between the two possibilities mentioned above entails a substantive difference in terms of the mental requirements each involves. In the former case, x has no idea of abstract standards and accepts others' behaviours as concrete standards for its own behaviour. In the latter, x knows that a standard (e.g., a norm) exists and may operate on Y. Here, we will focus on x' s search for the norms in force in the reference group.

• To find them, x must be able to infer norms from Y's behaviours. x must execute a series of interrelated tests; intuitively,
  1. The more frequent behaviours are good candidate for inferring norms; however, this criterion is not sufficient (and, as will be shown later in section 4, is not even necessary).
  2. A non-instrumental (and costly) behaviour may be interpreted as an obligation. For example, if a driver stops at the crossroad and no other car is in sight, the observer may infer that a "stop" sign is situated at the crossroad.
  3. Others' reactions to diverging behaviours are relevant indicators. Are others indifferent or do they react? And if so, how do they react? What are the consequences (costs) of the diverging action? Sometimes, especially if the consequences are very small, x might conclude that to take the norm into account is not worth the trouble.
  4. A good reputation of the conforming agents may suggest that a given behaviour is valued positively by the group. The conforming agents may exhibit it because it is valued positively (as opposed to out of utility or convenience) and in order to obtain a consequent positive evaluation from their fellows.
  5. Some effects of conformity may also indicate the norm's existence: the more a given behaviour has a positive effect at the global level (coordination), or meets existing values (e.g. female emancipation) or meta-norms (e.g. reciprocity), the more likely it is to be interpreted as a norm.
• To find the norm
  1. x must have a normative meta -belief, to be instantiated for a specific set of normative beliefs.
  2. x must also have a normative meta -goal: x wants to know what the norms are in order to (decide whether to) accept and possibly accept them. Beside, x must also have the belief that a norm in force among a given set of agents Y is generally accepted and observed by Y; x therefore has
  3. A normative belief about the efficacy of the norm (how much a given norm in force in a set of agents is applied by the members of the set; more precisely, the ratio between violation and fulfilment of the norm. Indeed, some norms might not be violated but also never fulfilled. In case of war one cannot use rusty swords. This norm will never be violated (anymore), but of course is also not applicable anymore. Efficacy is a quantifiable phenomenon, although under a minimal threshold a norm is probably (perceived as) no longer in force.
  4. Finally, x needs to have normative beliefs about what types of norms can be inferred by observing which sub-population (Y's subset).

Consequences

3.12

Mental consequences of social monitoring might include an update of the agent's values associated with given beliefs: for example, the agent may update the value of some beliefs concerning the frequency of given behaviours, the certainty of the related beliefs, and the diffusion or influence of given intentions and values^[5].

At group level

3.13

If distributed social monitoring leads to a reduction of the perceived discrepancies, it is most likely to serve as a means to:

• Enlarge the dimensions of the group and its sphere of influence and therefore
• Strengthen its power, amplify its visibility and effectiveness, and
• Justify/rationalise the maintenance/survival of the group
• Reduce its internal discrepancies, thereby
• Increase its consistency, and at the same time
• Emphasise its difference from other groups (its distinctive features).

All these features of a social group may prove advantageous for the members of the group in the long run, i.e. they may not profit from it directly, but may profit from the group in some future circumstances.

At the individual level

3.14

An important set of consequences concerns how x will use such norms once she has found them out. For example, x may apply them with no further need to monitor her social surroundings. Besides, x may generalise them to other contexts. More significantly, x may infer the rights and privileges that may be derived from them. Finally, x may go back to Y and evaluate its members with regard to those standards. The latter consequence leads us to what we will call norm-based social control.

Norm-based Social Control

4.1

We will speak here of norm-based social control as a type of social monitoring (see Figure 4). Agents exercise social control when they monitor one another with regard to a (set of) norm(s) that they have acknowledged as such, and with regard to which they want to reduce existing behavioural discrepancies. More precisely, x exercises social control:

Figure 4. Social control as norm-based monitoring

• When x has a normative belief about a group of agents Y and a norm n
• x accepts n
• x wants to monitor Y's behaviour with regard to n.
• x wants to reduce discrepancies between Y's behaviour and the norm n.

Note that norms may be formed through social monitoring, but also arise from other inputs (see Figure 4).

4.2

Again, some of the questions raised with regard to social monitoring apply to the case of norm-based social control:

• Which norms undergo social control? Do agents check one another with regard to any norm, or with regard to a subset of them, and if so which one?
• What are the motivations which induce agents to execute social control?
• Which minimal cognitive ingredients are required for agents to exercise (more or less effectively) social control?
• What are the consequences of such a phenomenon both at the global and at the local level?

4.3

The likelihood of social monitoring leading to norm-compliance is a function of different quantitative dimensions, which include the degree of efficacy of the norms observed.

Norms under Social Control

4.4

Here, we will restrict the application of social control to:

• A subset of norms in force in a given set of agents Y, or in a subset Y_i.
• x belongs to Y, but may not belong to Y_i
• x acknowledges n as a norm in force in Y, and relevant to Y_i.

4.5

This points to our analysis of norm acceptance (see Conte et al. 1998), where we modelled this process as consisting of a number of tests. One such test may give as an output that a norm is acknowledged as such (as issued by a legitimate authority, falling into the domain of competence of that authority). A subsequent test may give as an output that the acknowledged norm concerns x. In such a case x accepts n as a norm which concerns it.

4.6

In the case of social control, we hypothesise that agents are entitled and in fact do check one another with regard to acknowledged norms. Moreover, we hypothesise that this is the case even when agents are not directly concerned with the norm. So, a pedestrian will check the car drivers' compliance with the norms regulating the circulation of cars. In this case, driver and pedestrian have complementary roles: x is a beneficiary of the norms, and probably shares them, considering them as a valuable means for the welfare of the citizens. One might wonder what are the limits of entitlement to social control. Indeed, it may be argued that pure observers are not entitled to criticise the behaviours of society's members^[6]. A reasonable boundary is represented by the edge of the range of influence or strength of the authority (whether legal or social) that generates the norm. An observer of a religious sect cannot criticise the behaviours of its members with regard to the prescriptions of their religious authority. However, she can criticise their behaviours with regard to a shared institutional authority, even when such behaviour is not expected from her. As an atheist I am not entitled to check whether a catholic goes to church. But I am entitled to control whether he respects the precept of tolerance in force in our shared society with regard to other religions even if I am not affiliated to any.

4.7

One important dimension to take into account is the type of norms under consideration. Classifications of norms abound not only in the philosophy of law but also in social psychology (Cialdini et al. 1991). Here, it is proposed to make a distinction between descriptive, norms and personal norms. For reasons discussed elsewhere (see Conte & Castelfranchi 1999), we maintain that norms are prescriptions (whether legal or social), which might refer to a single agent, although generally they concern a multi agent system. In several cases, a norm may be posed by a given agent upon herself (as happens in commitment). Usually, this implies that a given meta-norm is (perceived by the agent to be) called into question and leads to the formation or explicit issuing of a specific norm. A norm is always a prescription more or less explicitly posed on a non-empty set of agents.

Local motivations

4.8

Why do agents check whether norms are complied with? There seem to be several types of motivation arising from the circumstances described above. The reader is warned against considering these motivations as mutually exclusive, since probably in natural agents they appear in various combinations:

• To benefit from the norm: if x is a beneficiary of the norm, she wants to assess the extent to which the norm is known/accepted/observed in a given situation
  1. Before deciding whether to join that situation: for example, x is a non-smoker invited to join a poker party.
  2. In order to put pressure on Y_i to observe n, or reduce its own damage (reparatory norm): for example, x is hungry and tries to convince the members of Y to observe the norm that states that hungry people should be fed.
• To share one's neighbourhood norms: x shares the norm and wants to be a member of Y_i only on condition that Y_i members observe it (this corresponds to the situation in which x is beneficiary, but here x' s ultimate goal is to share the norms with its social neighbours). E.g. x only wants to do research in a group that cooperates and shares all results of their research.
• To defend a shared norm:
  1. Know whether the shared norms are accomplished and, if they are not, have their efficacy increased, possibly by exercising some sort of normative influence (see section 5 and Figure 5).

     Figure 5. From control to influence

  2. Control the validity of the shared norms. Interestingly, the spread of compliance may be an indicator of the validity of the norm. If a norm is found to be ineffective, agents may want compliance to increase in order to reduce the "dissonance" between their attitude with regard to that norm and its poor efficacy (Festinger 1957). In such a case, normative influence and thereby normative control aim to reduce cognitive dissonance rather than increase compliance. For example, x tries to convince all members of Y (that are catholic) to attend church every Sunday.
• To maintain one's normative reputation: x believes that Y_i normative behaviour has an impact on the normative reputation of Y, which x belongs to and, in turn, upon her own. In some sense, x is an indirect beneficiary of the norm, or at least of its application.
• To control equity: x has complied with the norms and wants to check whether others do so or whether they are punished if they don't. Social comparison (distributive justice) leads agents to check whether others comply with the norms in order to check the resource distribution within the group. This should not be confused with a desire for equality and identification with others. Somewhat complementarily, it depends on a need for equity (cf. Homans 1974), and more precisely for normative equity (cf. Conte & Castelfranchi 1995, ch. 7). The general principle for equity states that agents want it to be the case that, benefits being perceived as equal, their own costs should not be (perceived as) higher than others'. A normative case of this principle follows: agents do not want their normative costs (the costs of normative actions) to be higher, benefits being equal, than those sustained by others. Ergo, agents check normative behaviours in order to compare their normative costs with those of others (see Figure 6). Of course, the follow-up question is the typical question which is posed by the equity principle, especially in its perceived form (cf. Berscheid and Wahlster 1978c): in which direction will agents decide to restore equity? Will they reduce their costs or act so as to increase those of others? A very trivial answer is based upon temporal considerations. Before complying, agents still have time to avoid inequity by either dropping their normative goals or by influencing others to comply. Afterwards, equity can only be restored through others: either by having transgressors punished or by influencing them to comply. Accordingly, when other agents have already transgressed against some norms, they must be punished for social balance to be restored. If they have not, there is still time to influence compliance. However, other factors beside time may intervene. For example, the degree to which the controlling agents share the norms may lead them to encourage norm compliance and enforce the norm or not. Perceived feasibility of norm influence may lead agents to prefer this option to punishing (e.g., denouncing) violators. Some pro-social values or norms (e.g., never betray peers) may conflict with the anti-social modality of restoring balance: depending upon the degree to which agents adopt the former they may be feel more or less inclined to resort to such a modality.

  Figure 6. Restore Equity

• To decide whether to comply with the norm. Agents check the efficacy of the norms in order to decide whether they themselves should (continue to) comply with them. One fundamental reason for this is that agents want to be like others (see section 3). But there may be other reasons:
  1. Infer the consequences of violation: x checks others' behaviour in order to infer whether it is safe to violate the norm: if others violate, surveillance must be insufficient or violation is tolerated.
  2. Obtain social approval (opportunistic compliance): if x' s normative goal is relative to a belief that the norm is effective in general or in given situations, x will accept the norm in order to obtain approval by other observers of the norm in question. For example, Jane knows what good manners prescribe at a dinner party, but she will keep to them only when others do so.
  3. Check the acceptability of violation: x checks others' behaviour when x wants to assess whether it is entitled or not to violate it. This is a consequence of equity, and somehow of the evaluation of the norm, but does not (only) restore equity. Rather, it provides x with social reasons, and justification for violation. For observers, poorly effective norms entail higher costs than an effective norm does (i.e., reacting to others' violation in case of one's compliance). As a consequence, one is more justified in violating a poorly effective norm than an effective one.
  4. Solve norm-conflicts: x checks others' behaviours with regard to a given norm when this conflicts with another norm and x needs inputs about how to solve the conflict. This actually goes back to social monitoring. In fact, what are the current principles for solving norm conflicts? How much is the norm hierarchy shared within Y_i?
• To fulfil a given responsibility with regard to the norm. x checks others' behaviours with regard to n when x plays a social role which prescribes the surveillance of Y_i and/or the efficacy of n, or when x is held responsible for others' violation. For example, x is in charge of writing on the blackboard the names of schoolmates who break the rule of silence (this leads to responsibility for the observance of the norm). Norm-responsibility is both an objective state and an individual motivation. It is objective, since an agent can be held responsible under certain conditions (cf. section 4), for the norms impinging on the group it belongs to, without it even being aware of this. However, agents may become aware of such a responsibility and act accordingly, whether they share the norms in question or not. A good example is that of a policeman. They might not actually share all norms imposed by law (like fining a pedestrian that crosses a red light), but they will still fulfil the norm, because of their responsibility as policemen. Interestingly, norm-responsibility may provide a justification for one of the previous motivations (often for resorting to an anti-social modality, like denouncing transgressors). Responsibility may be dropped or denied when agents do not feel like or are not sufficiently motivated to take any initiative into the direction of norm influence.(19)

Mental requirements

4.9

The motivations identified above require agents to be endowed with some further mental properties, namely

• x has normative beliefs about a norm n, including some belief about
  1. Effects of norm application on its own goals/interests
  2. Effects of norm compliance on the whole set of agents Y to which Y_i belongs, which in turn may be reflected upon x' s goals. (In the case of normative reputation, members of a group inherit the reputation of the group they are affiliated to.)
• x has a goal of normative reputation (i.e., x has the goal that the other agents in Y recognise that she always conforms to the norm or violates them.)
• x has a normative goal corresponding to the norm in question, but not yet the corresponding intention, and the process from a normative goal to a normative intention partly depends on the beliefs about efficacy of the norm and the certainty of punishment in case of violation.
• x has a responsibility for the application of the norm (see 4.10).
• x has a normative equity principle: x wants that others subject to the same norms sustain costs not lower than those she has sustained (see also 4.8).

Consequences of norm-based control

4.10

The consequences of social control at the group level are largely the same as discussed for social monitoring. The main difference is that the increased coherence of the group is based on a shared set of norms. Having explicit norms makes it easier to identify the similarities in behaviour for members of the group and also to identify the differences between the group and other groups.

4.11

A second possibility with norm-based control is to delegate the monitoring of certain norms to an individual member of the group.

4.12

At the individual level the consequences are more or less the same as already mentioned with social monitoring. The main new feature is that the possibility of social control will also lead some agents to exert normative influence on other agents.

Norm influence

5.1

Sometimes, agents decide to increase the probability that agents subject to a given norm will actually comply with it. Influence is therefore based upon control: agents will put pressure on others to comply with the norms or with a given norm, when they have reasons to believe that the norm efficacy does not exceed the confidence level. Again, there is no need for x to be a member of the set of agent Y_i upon which the norm is in force, but x must be a member of the set Y, to which Y_i belongs.

5.2

Norm-based influence is here defined as the normative goal that the norm be accepted and applied by Y_i. More precisely, x exercises a norm-based influence on Y when:

• x has a normative belief about Y_i and action a (that there is an obligation upon Y_i to do a)
• x wants that Y_i do a because Y_i has a normative goal about a;
• therefore, x wants that Y_i has a normative goal about a.

Motivations and requirements

5.3

Interestingly, even if x belongs to the subset Y_i, there is no need for x to be an observer of the norm. In fact, x may resort to normative influence whenever this is convenient for her, for example, when x is a direct or indirect beneficiary of the norm. For example, it is far from rare that people complaining about bad drivers' habits when they are pedestrians, exhibit exactly the same bad behaviour when driving their cars. The motivations are essentially a subset of those that we have examined with regard to control. Analogously, the mental requirements^[7] are a subset of the previous ones about norm-based control.

Consequences

5.4

The more the agents are sensible to intra-group influence, the more norm-influence is expected to lead to an increase of norm efficacy, for two distinct reasons:

• Norm-influence concerning one given (set of) norm(s) is likely to increase others' evaluation about the efficacy of that (set of) norm(s) (their monitoring) and this in turn will increase their compliance.
• Norm-influence will directly reinforce norm compliance, by different modalities:
  1. Discouraging transgressions: a typical example is offered by agents discouraging their neighbours from smoking; a sub-ideal case is provided by a driver turning on her car headlights to warn fellows coming from the opposite direction that speed control systems are nearby.
  2. Raising the social costs of disobedience (disapproval, denouncing, retaliation).

5.5

When, to what extent and in which modality is norm-based influence effective? And since norm efficacy, as was shown earlier in this paper, has an impact on agents' normative beliefs, what is the effect of norm efficacy on norm-based influence? Computer simulations could be designed to check the interactive effect of norm-based influence and beliefs about norm efficacy on the propagation of different types of norms and conventions in multi-agent systems.

The necessity of more complex agents

6.1

Social monitoring, as discussed in this paper, is a rather complex social capacity of intelligent autonomous agents. It requires a more complex agent model and architecture than is usually implemented. What are the advantageous of such a view of social monitoring, which cannot be obtained by means of simpler instruments, for example with adaptive and learning systems?

6.2

This is not the forum for re-examining available adaptive learning systems. However, we essentially aim to draw the readers' attention to two complementary sides of social monitoring, imitation and social control. While imitation is a very popular notion (although often oversimplified and poorly or inadequately treated), the complementary aspect of social control is usually ignored or under-estimated within the computational fields of science. Indeed, it is considered as a puzzle within other fields, such as sociology. Our claim is that imitation and social control are interrelated aspects of one common phenomenon, i.e. social monitoring. By understanding this common ground, it is possible to see that agents check others' behaviours on the grounds of pre-existing internal criteria (for example, norms). Consequently, the emergence of social norms and standards, far from resulting from mere mutual adaptation, is a complex social and cognitive process. In this process, agents compare their own criteria with those of others, and extract from others' behaviours cues allowing them to specify abstract normative precepts.

6.3

What is the advantage of modelling such complex process, over the simpler, behavioural approach to social learning? There are several advantages:

• An explicit model of social monitoring allows to account for social control: how could one explain social control otherwise (without resorting to the somewhat convoluted hypothesis that agents control others to obtain their approval)?
• It also allows the interrelationship between imitation and social control to be modelled explicitly: agents are both actors and targets of social influence. And, what is more, they often exercise both roles at the same time! I may turn to you to appreciate how efficacious a given norm is, and therefore to control your acceptance of the norm, but if I observe that you violate the norm, I may conclude that this norm is not to be respected and may decide to conform to your behaviour.
• At the same time, it allows one to model social and legal norms as original sources of influence, and criteria for social comparison. This in turn has some more specific advantages
  1. Account for social and institutional innovation without calling necessarily into question evolutionary mechanisms.
  2. Perceive the similarities between social and legal norms, which are often considered to be totally separate.
• Finally, an explicit model of social monitoring allows the forerunners of social control and normative influence to be explored. Norms represent but a special type of criterion in the process of social monitoring. But the scope of this process is wider, and concerns also non-normative domains. As social psychologists perceive, agents check one another in many different contexts to achieve many different goals. Consequently, the range of applications of this activity exceeds the mechanisms of social order. Thanks to social monitoring, agents can achieve a more efficient preventive cooperation in teamwork, avoiding errors that might be irreparable or the costs of repair.

6.4

Essentially, all these results imply the capacity to compare others' behaviours to one's own internal criteria. None of them can be achieved if agents were enabled to simply adapt to one another's behaviours.

Research agenda

7.1

In this paper, we have formulated the requisites of a model of social monitoring. A great deal of work needs to be done in order to translate such a model into an agent platform, and implement it in order to obtain results in some relevant domain. Mainly, urgent developments should concern

• Specify the conditions and limits of procedures: when do agents monitor one another? When does social monitoring lead to norm-based social control? When do agents drop their normative beliefs and adopt others'? Under which conditions and to what extent do agents accept others' inputs in the formation of normative beliefs? We have sketched a number of potential tests, but the conditions for their applicability need to be clearly specified.
• Procedures ought to be implemented in a well-defined domain of application, with regard to a specific task or problem. These ought to be formulated both within a normative and a non-normative domain.
• Possibly, the present approach ought to be compared with a behavioural approach by means of simulation. When, in which context and for which effects do adaptive agents achieve satisfactory behavioural standards at the global level? Here, a number of specific hypotheses ought to be formulated and explored by means of simulation. For example, a behavioural approach probably allows fast convergence on given standards. But what about emergencies? Which approach is preferable when innovation is necessary in order to deal with urgent environmental demands? Analogously, what about intelligent retrieval? In other words, which approach allow agents to drop a given norm in conditions of material impossibility, or emergency, or (norm) inconsistency, and restore it when conditions normalise?

Conclusions

8.1

To understand how agents should be modelled in order to act in a multi-agent system (or society), we must understand the social rules that govern the behaviour of the individual agents. The social rules should be such that they strike a balance between the individual autonomy of the agents and the effective functioning of the complete society. In human societies, the social rules are formed through a complex set of social norms and conventions. In order to know which parts of this complex social system are of use for multi-agent systems we have to study the concepts which are important for human societies.

8.2

In this paper we have sketched a first analysis of the notions of social monitoring and control. We have put a particular emphasis on norm-based monitoring as we think that it is of prime importance for multi-agent systems as well.

8.3

In 4.9 we touched upon the notion of responsibility. We indicated why this notion is central for any complete account of social control and influence. A formal account of this notion is left for future work, which will also form a basis for a more formal description of the processes of social monitoring and control.

8.4

Future studies should be aimed at:

• Providing a formal model of the notions proposed here in terms of the mental attitudes they involve, and of the process leading from one to the other in terms of the mental rules and operations accomplished on them;
• Designing a set of computer simulation experiments for checking the role of norm-based control and influence on the propagation of norms and conventions in multi-agent systems.

Acknowledgements

We acknowledge with thanks the contributions of the European Union Framework Projects ALFEBIITE and FIRMA.

Notes

¹ Throughout the paper, we will speak of 'agents' to refer both to artificial, simulated agents, and to natural, real-world systems. The term 'actor', used by social scientists to refer to social systems especially at the individual level, evokes the script-based view of social action, which is not consistent with the conceptualisation of the present authors.

² Note that by normative agents we do not refer to a characteristic of personality. In our meaning, a normative agent does not necessarily observe the norms, but is able to decide whether to do so.

³ In terms of (1) language and formalisms used (strictly logic-based in the logical-philosophical domain and more oriented to implementation languages in the (multi-)agent domain); (2) theory of reference (philosophy of law and deontic philosophy in the former domain as opposed to agent theory, game theory, AI in the latter); (3) objectives (expert legal systems, theory of institutions, in the former case as opposed to social theory and optimisation of coordination and cooperation in the latter) (cf. Conte et al., 1999).

⁴ This is shown by the example of looking for a shelter from rain. By observing what happens to another agent, the observer learns to avoid trees, where a negative event (lightning) is likely to occur. Here, the observer learns a negative effect of a known plan of action.

⁵ A caveat is necessary. What are the relationships, the connections among social monitoring and other related phenomena, for example, social comparison, learning, task execution monitoring? Social monitoring represents a means that under different conditions (goals and beliefs) may lead to different outputs. For example, a perceived discrepancy of powers may activate the mechanisms, rules and principles associated with social comparison (the equity principle), or may lead to updating the set of beliefs concerning one's position in the social hierarchy. A perceived discrepancy of behaviours may activate social learning. Given our main interest in the issue of norms spreading and the role of normative intelligence in such a phenomenon, we will examine these questions considering the contributions that a model of social monitoring can provide to clarify the issue of social control.

⁶ We are indebted to an anonymous reviewer of a draft of this paper for this interesting remark.

⁷ An interesting question here is what x can do to achieve normative influence. At least three modalities occur:

• Prevent transgression
• Promote obedience (through persuasion or reinforcement)
• React to current violation to promote obedience (through disapproval and reproach, denouncing, isolation, retaliation)

Further analysis should be carried out in order to strengthen and enrich this typology. Moreover, computer simulations of normative agents in interaction could be designed to check the respective effects of these modalities.

References

BANDURA A (1977) Social learning theory. Englewood Cliffs, NJ: Prentice Hall.

BERSCHEID E and Wahlster E (1978) Interpersonal attraction. Reading, MA: Addison Wesley.

CHAVEZ A, Dreilinger D, Guttman R and Maes P (1997) "A real-life Experiment in Creating an Agent Marketplace". In Proc. of the 2nd International Conference on the Practical Applications of Intelligent Agents and Multi-Agent Technology (PAAM'97).

CIALDINI R B, Kaligren C A and Reno R R (1991) "A focus theory of normative conduct: A theoretical refinement and reevaluation of the role of norms in human behaviour". In M.N. Zanna (Ed) Advances of Experimental Social Psychology, San Diego, CA: Academic Press, 24: 201-234.

CONTE R and Castelfranchi C (1995) Cognitive and social action. London: UCL Press.

CONTE R and Castelfranchi C (1999) "From Conventions to Prescriptions. Towards a Integrated View of Norms". AI&Law, 7: 323-340.

CONTE R, Castelfranchi C, Dignum F (1998) "Autonomous Norm Acceptance". In J. Mueller (Ed) Proceedings of the 5th International workshop on Agent Theories Architectures and Languages, Berlin: Springer.

CONTE R, Falcone R and Sartor G (1999) "Agents and Norms: How to fill the gap?" Artificial Intelligence and Law, 7 (1):1-15.

CONTE R and Paolucci M (2001) "Intelligent Social Learning". Journal of Artificial Societies and Social Simulation, 4 (1): https://www.jasss.org/4/1/3.html

DIGNUM F and Conte R (1997) "Intentional agents and goal formation". In M. Singh et al. (Eds) Proceedings of the 4th International workshop on Agent Theories Architectures and Languages: Providence, USA.

FESTINGER L (1954) "A theory of social comparison processes". Human relations, 7: 117-140.

FESTINGER L (1957) A theory of cognitive dissonance, Stanford, CA: Stanford University Press.

HART H A (1961) The concept of law. London: Oxford University Press.

HOMANS G C (1951) The human group. N.Y.: Harcourt.

HOMANS G C (1974) Social Behaviour. Its elementary forms. N.Y.: Harcourt.

KAMINKA G A and Tambe M (1998) "What's Wrong With Us? Improving Robustness through Social Diagnosis". Proc. of the National Conference on Artificial Intelligence (AAAI-98).

LIEBRAND W. and Messick D (1996) (Eds) Frontiers of social dilemmas theory. Berlin, Springer.

OSTROM E (1998) "A Behavioral Approach to the Rational-Choice Theory of Collective Action". American Political Science Review, 92(1): 1-22.

SHOHAM Y and Tennenholtz M (1995) "Social Laws for Artificial Agent Societies: Off-line Design". Artificial Intelligence 73 (1-2): 231-252.

SHOHAM Y and Tennenholtz M (1997) "On the Emergence of Social Conventions: Modeling, Analysis, and Simulations". Journal of Artificial Intelligence, 94(1-2): 139-166.

TAMBE M (1997) Towards flexible teamwork. Journal of Artificial Intelligence Research, 7: 83-124.

WALKER A and Wooldridge M (1995) "Understanding the Emergence of Conventions in Multi-agent Systems". In Proceedings of ICMAS (International Joint Conference on Multi Agent Systems) (San Francisco).

WEIBULL J W (1996) Evolutionary Game Theory. Cambridge, MA: The MIT Press.

Return to Contents of this issue

© Copyright Journal of Artificial Societies and Social Simulation, 2001