As states Robert Axelrod [1], "agentbased modeling is not only a valuable technique for exploring models that are not mathematically tractable; it is also a wonderful way to study problems that bridge disciplinary boundaries". Some earlier developments of multiagent systems as an emerging software paradigm in the domain of artificial intelligence are closely related with Marvin Minsky ideas about the "mind as society" [2], so that it is easy to find a kind of social bias in any MAS. Conversely, earlier developments of social theory or social philosophy try to make some kind of formalization or modeling, not just empirically descriptive but "building ideal societ ies" in order to explore its consequences or to make experimental "control comparis ons" with the real ones. Part of the works of Plato, Moro, Campanella, Bacon, Fourri er and other classic utopian social thinkers could loosely illustrate it, but the socio economic developments around the formalization tradition and the analytical soci ology can better illustrate this idea [3] [4]. There are multiagent systems, with a so cial flavor, and there are social theory, with a formalization flavor. Then, agentbased social simulation (ABSS) can be seen as a clear connector or bridge between at least two disciplinary boundaries: those corresponding to some of the social scientists' community and some of the computer engineering's community.

The list of social processes and phenomena subject to research by mean of formal ized methods, modeling and experimental computer simulating is large and is increas ing as the JASSS index can show [5]. Some recent books about general sociological theory incorporate results of ABSS research [6]. But, in spite of the "clear bridge" and the growing interest, actually it can be said that the ABSS or other kind of simula tion approaches are far from belong to the mainstream of social sciences research.

Our argument is that we cope with a "Social Scientist Dilemma": without a strong programming competence the social scientist finds hard to build even the most simple artificial society, but the social scientist is forced to necessarily consider the high complexity of any social system. To solve this dilemma, for each research project, an ABSS researcher have to balance the amount of complexity that she/he wants to im plement into the model against the programming learningcurve effort needed for that level of complexity.

2. Theoretical Vs. Empirical Quality and the two Communities.

Research is not an individual issue, but an institutional social practice, so the "Satisfa cing Tradeoff" criteria for the evaluation of a social simulation model are a matter of social scientists community. Although there exist some efforts to establish quality standards for ABSS research and results dissemination [7] [8] there are still no clear, distinct and common guidelines available. This is a relevant issue, because the sci entific acceptability of an ABSS research depends on the mentioned dilemma. 2The question about the meaning of satisficing tradeoff for social scientists leads to the problem of setting common criteria to evaluate the quality of research. A satisfa cing simulation research in any domain can 1) provide excellent quality in the simu lated output data, fitting with actual empirical information ("empirical quality" EQ), and can 2) be highly coherent with the set of theoretical most accepted and updated knowledge for the specific domain both in the "fine grain" description of the model, in the rules governing the system evolution throughout the time and in the initial ad justments of the simulation ("theoretical quality" TQ). These two cross dimensions provide four ideal types as show Table 1. A plausible conjecture about the order or preferences among ABM practitioners is that there is a significant degree of heterogeneity between different communities. So, at least two variants could exist: an academic/research variant (ex>vn>bb>in) and an engineering/professional variant (ex>bb>vn>in). For instance, in regard to computa tional economy, López has enunciated this distinction as follows: "From an engin eering approach it is pursued to generate ideal entities that could act in actual mar kets. From a social sciences approach it is intended to replicate the most realistic agents attainable taking part in artificial markets, in order to understand how the transactions among actual individuals get organized". [9] There is no systematic research developed until now to confirm the general hypo thesis about the heterogeneity in the arrangement of preferences, as well as to estab lish the specific characterization, for different communities of ABM practitioners. With no need to formalizing here this argument by means of Game Theory it can be said that the actual existence of the aforementioned heterogeneity between communit ies can be one of the motives for some problems that have an effect on multidisciplin ary collaborations in the domain of applied simulation, that is to say, between aca demic specialists in social sciences and engineering specialists in computer simula tion. It is necessary to choose a variant in order to face the problem of social systems complexity. From an understanding perspective, so that the analysis of social pro cesses and systems attempt to produce an organized and coherent set of theoretical as sumptions about the mechanisms that generate the observational data outputs 3 , is preferable to assure high levels of theoretical quality (ex, vn).

The choose of this variant imply that any option lowering the theoretical quality in order to assure empirical data fitting must be rejected as a mere secondbest, for in stance, any knowledge approaching by means of automatic building of neural net works. In other words, it can not be considered satisficing those models and simula tions that generate datasets with a high adjustment with empirical observations but lacking of "fine grain descriptions" about the system elements and relationships as for the actual state of theoretical corpus of social knowledge.

Of course in any model a "fine grain description" is a formal representation of the object system, and therefore, more or less, a simplification. But, simplification does not necessarily means "black boxes". Moss & Edmonds point up to a pair of AMSS properties that should attract the interest of sociologists [10]: 1) they capture features of actual social order producing data with empirical relevance, but also 2) they natur ally draw upon and cohere with "detailed" core strands of sociological literature. The same idea can be traced in Epstein's research programme on generative social science and other wellknow epistemological proposals. [6] [12] 3. Levels of Social Complexity: Necessary in an ABSS?

So, ¿what will be the best approaching to complexity in the social domain by means of ABSS? In accordance with the preceding arguments the answer must be: including into the simulation model all the relevant elements in line with the most updated the oretical knowledge, even 1) if it must be shaped as plain simplifications and 2) if it will not produce accurate datasets fitting with empirical observation. This second is sue is a main research field in ABSS and there is relevant work about the validation problems [11]. As for the first issue, if social scientists community considers "com plexity" as a key issue then the first step is to inquire into the sources or complexity that relates with ABSS. It is not just the heterogeneity of the human behavior con sidered as result, but especially the heterogeneity of 1) the basic elements that gener ate such results, and 2) the diversity of levels in which such basic elements interact.

At least five levels of complexity4 can be recognized in any ABSS of a complex so cial system5 :

• 1level: The basic complexity of the cognitive "subsystem" of each social agent, which implies a more or less complex cognitive model of environment perception and (re)action. • 2level: The added complexity of social agents with operating and evolving "social images/maps" about other's behavior expectations (i.e., trust, social norms, institu tional order...). • 3level: The added complexity that becomes from the heterogeneity of multiagent interaction contexts, both considering some kind of typical situations such as Game Theory formalizations [4] [12], or some kind of sociability contexts such as Rela tional Models [13]. • 4level: The added complexity of multiple individual actions aggregation, which dynamically generate patterns of macrosocial outcomes, commonly know as "first order emergence". • 5level: The added complexity of loopback systems, with effect processes from the macrosocial into the most basic levels by means of "cognitive reconfiguration" and "experience learning", or "secondlevel emergence" [11].

Each of these levels affects the "upper" one, except for the last loopback, specially linked to 1level and 2level 6 . The 0level does not apply to actual social systems, but to any computed simulated society. In the 1level the complexity is shared with any other ecological system; the 2level and 3level can be considered human or DAI specific. The 4level and 5level complexity is shared with any material multi particles system. Although most of the complexity levels can be found in nonsocial systems, some of the mechanisms that rule the evolution of social systems over time are human specific, for instance, the social labeling learning.

While social simulations are just simplified models of object social systems or parts of social systems, is not necessary to include every complexity level the implementa tion. Rather, the specification in detail of every level of social systems complexity into a ABSS could result in a computational trap (0level). Of course, every simula tion is build on the basis of a particular research problem, and that is the reason why it is not needed to include every level. But, if the case is about establishing the minim um complexity level for a generic simulation builder or generator, then the criteria should be to include all relevant levels of complexity that assure the wide community acceptance on the basis of the most updated theoretical foundations into the social do main.

Regardless of this general theoreticalincluding criterion, there are a number of theoretical models that have been not yet included into any computer simulation, so that they are not validated. A simple methodology to validate or dismiss hypothesis about the relevance of a certain complexity nlevel or about some element or mech anism is to check the outcomes from a simulation that includes it against an identical simulation except for the lack of it. Then, if data analysis supports the null hypothesis of no significance differences between outcomes we can conclude that the nlevel (or element or mechanism) is irrelevant. In ABSS, due to the 0level complexity, it is no acceptable to include any irrelevant element, because of the risk of decreasing compu tational performance. Following this methodology, after the implementation of an ad ditional feature into a base simulation model, the further the variant simulation out comes deviate from the base simulation outcomes, the greater the doubt cast upon its validity as a relevant element.

As a plain example of this methodology, hereafter are presented some guidelines for checking a theoretical model "located" into the 3level of complexity, that is, the social complexity that becomes from the fact that social agents interacts with each other in a different way. For instance, in terms of coordinate an action a group "can seek a consensus of the group as a whole, the chief can decide (and delegate minor aspects of the decision), people can vote, or they can use a market mechanism based on utility or prices" (Fiske, 2005). There are many ways to solve each interaction context, but the Relational Models Theory (RMT) posits that "human relationships and social systems are culturespecific implementations of just four elementary rela tional models in various combinations" [17][18] [19]. Concerning interchange of goods, services or information 7 , this four RM can be characterized as follows:

• Communal Sharing (CS): An equivalence relation, so that agents in each group are the same in respect to resources, so they share each other but not with outsiders. • Authority Ranking (AR): A linear hierarchy in which agents are asymmetrically differentiated, so that the "upper" agent can take resources from the "lower" one, but not the opposite way. • Equality Matching (EM): The agents keep track of the additive differences regard ing the interaction partner resources, with an even balance as the reference point. • Market Price (MP): A relation based on a socially meaningful proportionality, so that the interchange will be ruled by a consensus ratio (monetary unit, utility, effi ciency, effort, merit, or anything else).

This idea of a small and clearly specified set of relational models or "schemata" [20] is a good candidate to be considered one of the essential pieces of social theoret ical development to be included in an ABSS builder. But, this simple idea will make simulation more difficult with regard to the 0level. This is a "Satisficing tradeoff Dilemma" situation, as it has been described previously, and a validation or dismiss checking will be of maximum interest.

An existing simulation model can be chose to check against it the "addition" of this level of complexity as control model. The requirement for the basic model is that the agents should have some cognitive capabilities to percept the environment, some social mapping capabilities, can perform actions that affect the environment and other agents, and can reconfigure (learning) his own cognitive model as a loopback effect of the system macrostates. That is to say, the basic model performs in such a way that considers the 1, 2, 4 and 5 complexity levels. An additional requirement is that exist some kind of performance indicator to be tested against alternative simulation models.

A slightly modified version of the very wellknow "Shopping Agents" Netlogo mod el8 from Gilbert & Troitzsch [21] should be useful as a base model. In this model there are 10 shopping agents and 12 shopobjects scattered over a torustopology artificial world. Each shop has and endless stock of one product and each shopper have a shop pinglist of 10 different products to buy. The shoppers have to go to all the shops that sell the product in the shoppinglist until this list becomes empty. Each time turn (tick) the entire set of shopping agents move around the world, and can buy product if they reach a patch where a shop is located. Shopping agents can build its own data base of shop locations ("memory") in two ways: they can remember the location of shops where they have been, and they can interchange this information with other shoppers when meet another agent in the same patch. The movement of each shopper are a function of their "state of mind", that is, the shoppinglist (as motivational goals) and the memory (as an environmental knowledge database), so that shoppers move towards the location of shops whose product are part of the shoppinglist or move at random if can not reach any goal from the current memory. The simulation stops when there is no shopper with nonempty shoppinglist. The performance indicator of the simulated system is the number of ticks until reaching the stop condition, in Gil bert & Troitzsch words, "how long it takes them to complete their shopping trips" ([21]: 182).

In the interaction complexity 3level, the agents of this basic model acts using the Communal Sharing (CS) model, or schema, as there was just one group. In case of a concurrence of a pair of shoppers in a same location, each one always shares all the shops information to each other by means of an unconditional full information sharing procedure. In the interaction complexity 2level, the agents of the basic model have no record about the other agents, even after a sharing information interaction with a partner so, even if they have an environmental memory of shop locations, they lack of any kind of social mapping or image. Some modifications must be done in order to introduce the 2level and 3level of social complexity by means of Fiske's relational models of sociality. First, the proper operation of RM implies a number of new agent attributes to be modeled. First, per ception and recognition of other agent's relevant attributes stands on individual ex ternal observable features. So there is a requirement to give each agent a new attrib ute, different from the Netlogo identification builtin variable "who", that can be per ceived by other agents and used to apply the respective relational model. A chain of digits can be assigned to each shopping agent as external observable features. This could be understood as the agents "chromosome" [22] [21], and in a later extended version of the model can be used in the offspring recombination process to study some evolutionary properties. In a straightforward approach, a random number in the 110000 range will be sufficient.

Second, the proper operation of RM implies a number of new procedures to be coded. At least a social recognition procedure must be included, so that agents can build, along any simulation running, a list of other agents that they previously meet. Similar to the previous visited shops list, this social memory keeps track of the inter action record and could be used in a later extended version of the model into a more complex "social labeling" procedure that helps agents to make decisions about co operation with partners. So a new attribute must be included in the shopper's creation procedure to complete the basic shopmemory with the added socialmemory.

Next, there are tree kinds of knowledge necessary for the agents to be able to oper ate RM with social competency. These are the cultural features that children, immig rants and sociologists/anthropologists must learn to "enter" into any social com munity:

• Competence to recognize the relevant individual attributes, as group pertinence (Communal Sharing), social rank (Authority Ranking), balance criteria (Equality Matching) and proportionality interchange unit (Market Price). • Competence to recognize where each RM operate, into a variety of cultural mean ingful interaction contexts, as for instance the workplace, the family at home, the family with outsiders, a queue, an emergency situation, or a casual meeting. • Competence to correctly operate each RM, as rules or criteria to make a decision about the exact amount of cooperation to give.

To proceed step by step, version 2.1 will implement the necessary requirements to deal with Communal Sharing model of interaction and then validation checks will be done about different setup conditions of CS model against the basic model. Next ver sions will add other relational models with the same methodology. The final version include agents that can perform any four RM and a set of metarules to decide which one to apply in each specific dyadic interaction. This metarules are strongly environ mental or contextdependent and can be considered as a model of "cultural traits".

The requirements for the CS model to be checked are: 1) the possibility of setup the model with a different number of groups, and with a different proportion of agents of each group, 2) the shopping model characterizes a certain social context, so there is no need to open the model to other social contexts, and 3) the basic shopping model does already include the rule of information interchange for CS model, so there is no need to add new procedures but just to establish a kind of group filter that should trig ger or refrain the "full information sharing" procedure as a function of the interaction partner recognition. The recognition procedure can use a narrow version of "labeling", where simple procedural rules affect the agent action at every partner meet without memory, or a complex version of two phase "labeling" procedure that rules the updating of a social mapping or memory: a) "substantive labeling" at first meet based on preos rules (following Fiske [13]: 281), and "procedural labeling" after experience based on interaction outcomes evaluation [23] [24].

The general requirements for the AR model are the same than for CS model, but there it is needed to modify the information interchange procedures to establish a uni directional transference from the lowerrank agent to the upperrank one. The trigger ing of this new cooperation procedure is a function of the interaction partner recogni tion of relative social rank. The use of a social rank recognition procedure can give rise to relevant ontology matching problems [25] [26], that could be solved in a two fold manner: 1) to assume a centralized hierarchy ranking, with no error possibility may be coded into the agents observable "chromosome" like chevrons, and 2) to build new procedures for ontology alignment, or even let the agents interact under discordance consideration of the partner's relative rank what could be a source of conflict.

Finally, the requirements for the EM and MP models are not yet completely estab lished up to the present, but preliminary analysis give some clues and the suitable models are under active development. The following sections will show some early results, after a brief description of the quasiexperimental setups. 9 6. Some quasiexperimental results about checking Fiske RM. 10 Experiment #1: "Leave the market Vs. Keep chatting" Global performance.

The proposed methodology prescribe to experiment with model features that can generate relevant differences between outputs, so that it can be checked, for instance, the relevance of "leave" the market once the agents complete their shopping against the base model, where agents wait in shops and continue providing their location knowledge to other agents.

Figure 1 shows the distribution of outcomes for the "leave market" variant, and Figure 2 shows the basic model outcomes (for 1000 runs each model). With a tvalue of 1.4082 for a parametric two means test, and a p(t) = 0.1592 (>0.05) it can not be said, with a 95% of confidence, that there exists significant differences between the outcomes. So the variant seems to be irrelevant in this context of Communal Sharing relational interaction 11 .

The global performance of a single task could not be the best indicator of aggreg ate behaviour in a social complex system. ABSS can generate recordings of interac tion patterns between agents e.g., the evolution of "meetings", "recognitions" and "informationtransfer" against time. In this context, "meeting" means the simultan eous concurrence of a pair of agents over the same patch, "recognition" means a first time meeting between a pair of agents so that each one update its own socialmap by recording the distinctive features of the other one, and "informationtransfer" (IT) means the basic "talk" process between agents that brings to transfer information about the shops location to other agent in a meeting.

Figure 3 displays the typical shape of some interaction evolution indicators in a "keep chatting" model, while Figure 4 shows the typical pattern in a "leave market" model 12 . Although the global performances are quite similar, there are a salient qual itative differences in the interaction patterns along the simulation time. If shoppers re main in the market after they complete the shopping, they exponentially increase the probability of relevant information transference, as shows the high step curve on Fig ure 3. If shoppers leave the market, the probability of social information transfer will decrease over time, as shows the Sshaped curve on Figure 4.

Even if global performance quantitative indicators can be very useful for testing of alternative models, careful attention must be focused on the qualitative traits of the simulation time evolution outcome data because similar global performance can be achieved by means of a range of different interaction patterns. Experiment 2 proves how complex social systems (actual or simulated) must be analysed in the "fine grain" of overtime interaction, and how ABSS can be a suitable tool to do it.

Up to now, the experiments has been performed on a simulation model that imple ments a CS Relational Model with one social group: that is to say, all agents should interchange its own complete set of information, about shops locations, whenever they meet any other agent. Prior to testing the performance of another behavioural models (i.e., CS vs. AR), experiment 3 can help to explore the effect of the number of "socially relevant groups" into a CS system a sensibility analysis to validate ABSS.

Group membership is relevant in a CS system because, after a meeting on the same patch, the recognition of other shopping agent as insider or outsider will trigger (or in hibit) the IT procedure bidirectional informationtransfer of shops locations.

Figure 5 shows four distributions of global performance, using the simulated time (ticks) until the last shopper buy its last product. A replication of 1000 runs for each case provides the presented histograms of frequencies. The upperleft case (g=1) is the basic or control model; in the other cases there was different number of groups. The distribution of frequencies for the final time shows how the increase in the num ber of groups is related to the increase in time to finish the shopping simulation. That is because of the decreased probability to meet another agent suitable for "informa tiontransfer" in a CS context.

With a Fvalue of 847 for a parametric ANOVA test, and a p(F) = 0 (<0.05) it can be said, with a 95% of confidence, that there exists significant differences between the four models. So the number of groups, in a social system ruled by the CS behavi oural schema, seems to be relevant because of the social network topology that will be implied. But, again, these quantitative global findings must deserve some "fain grain", or qualitative, attention.

Figure 6 shows four typical final ITnetworks from simulation runs in social sys tems ruled by CS behaviour. Agents can be distinguished by its external traits, been the first number of each agent code (or chromosome) the main source to group alloca tion. The distribution of agents into different group follows a random function.

For this particular run, the time each agent spends to complete the shopping ranges from 852 ticks (shopper 17381) to 3962 ticks (shopper 16774). But if we take into ac count the intragroup means of ticks, that is 2161.3 for group1 (n=4), 3256.5 for group2 (n=2), 2526.7 for group3 (n=3) and 3283 for group4 (shopper 43719), a sound hypothesis is that, in a context of CS behaviour and many groups, any perform ance depending on social information is a direct function of the number of members for each group. Large groups performs better than small ones.

A replication of 4000 simulation runs, divided in CS models with 1, 2, 3 and 4 groups of approximately equal membership agents (Table 2), can provide some sup port to the previous hypothesis, although further analysis is needed.

Experiment #5: "AR Vs. CS Model, and number of groups" Global performance.

After developing a new artificial society where agents behaviour is ruled by the Authority Ranking model (AR), Figure 7 shows four distributions of global perform ance. In the upperleft case (g=1) there was only one group: the basic or control mod el, where AR rule works like CS rule.

With a conventional ranks hierarchy used commonly by all agents, the AR rule about transferring information has been implemented as follows: when they meet in the same patch, agents give all its own shops location information to other agents with equal or higher rank. Checked versus the corresponding CS models, the global per formance of AR models shows significance differences (Table 3). The overall per formance or AR is better in each case, and seems to be an opposite function of the number of groups.

As regards to the effect of the number of groups over the global performance in AR artificial societies, with a Fvalue of 295,97 for a parametric ANOVA test, and a p(F) = 0 (<0.05) it can be said, with a 95% of confidence, that there exists significant differences between the four models. The distribution of frequencies for the AR final time (Figure 7) shows how the increase in the number of groups is related to the in crease in time to finish the shopping simulation. Like in CS models, is it because of the decreasing probability to meet another agent suitable for "informationtransfer" in a AR context ?.

Results for experiment 4 could support the hypothesis that, in a multigroup CS context, large groups performs better than small ones (due to the corresponding high density network). Figure 8 displays four typical informationtransfer networks, as it arises from simulation runs in social systems ruled by AR behaviour.

On the contrary of CS context, in AR ruled artificial societies, the increasing of the number of groups will lead to high density networks (Figure 8: AR g=4 vs. Figure 6: CS g=4). This emerging topology could help to understand the global better perform ance of AR compared with CS, but appropriate understanding of the output data im plies to identify the different mechanism effects of "exclusive sharing" (CS) and "lowerrank exploitation" (AR) over the topology emergence.

Taking into account the intragroup means of ticks, for a single typical run, that is 1827 for group1 (n=3), 2208 for group2 (n=2), 2122.5 for group3 (n=2) and 1221.7 for group4 (n=3), seems to support the hypothesis that group performance is a direct function of the group size, an effect that overcomes the rank position effect.

A replication of 4000 simulation runs of AR models with 1, 2, 3 and 4 groups of approximately equal membership agents (Table 4) can provide comparison elements with CS group performance (Table 2). The corresponding ttest for each pair of group performance means shows significance differences between every CS and AR groups (exception g=1). So it can support the hypothesis that AR and CS models are different in group performance, not just in global performance.

A comparison between Tables 4 and 2 will support the conclusion that the group performance is a direct function of the rankposition (G1<G2<G3<G4). Although fur ther analysis is needed to discard the group size effect, it can be said that in a CS con text "cooperation" (less groups, or larger group) performs better, and in a AR context more "exploitation" (more groups) performs better.

In this presentation, we studied the problem of tradeoff between including theoretic ally required elements against excluding irrelevant levels of complexity, a dilemma that any ABSS practitioner must cope with. It is been argued that there exists hetero geneity between ABSS communities, and this has been characterized, in general terms, by the ordered set of preferences of the two main variants. Therefore, follow ing the academic variant, we make some remarks about the acceptability of a certain social simulation among the social sciences community. A first result is that ABSS methodology could not be easily accepted by social science community if it fails to fulfil two kind of requirements, been the first to include as much social systems com plexity and the second one not to increase the "technical" level of complexity for gen erate the artificial societies under consideration (0level).

Concerning these two requirements, here we have presented, by means of a case exemplification, a methodology to check the relevance of any theoretical element that could be considered as a candidate to be included into an acceptable social simulation. Our approach tries to test the relevance of any model by analysing the outcomes as if they come from experimental data. By generating pairs of simulation models that dif fers just in the issue under consideration, some theoretical elements could be dismiss from been a necessary part of any social simulation if the data generated by each model supports the null hypothesis of irrelevance. This methodology could be mis taken for standard ABSS sensitivity analysis, but there is a clear difference: in sensit ivity analysis the setup space been explored corresponds to a single model, but in the checking methodology the contrast explored refers to a pair of models the control and objective.

Experiments 3 and 5 seems to show that the global performance of the shopper's social system is a function of the number of groups (g), and some results from experi ments 4 and 6 supports the idea that in a context of many groups, the average per formance for a group depends on the extension of membership. In the two coordina tion models here considered, both CS and AR, as g increases, the global performance decreases, but some groups will perform better: in a CS model the large groups, and in the AR model the higherrank ones.

All this experiments and conclusions aims to exemplify the suitability of the use of ABSS simulation models to test theoretical claims, as a general methodology in the social domain theoretical research. Standard statistical quantitative tools can be com bined with a quasiexperimental approach (due to the simulated nature of the output data) to discard non relevant theoretical proposals. The microlevel data generated by a ABSS adds the opportunity to use some qualitative analysis tools that can bring "fine grain" understanding of some microfundamental mechanisms operating in the case study been considered.

Up to now we have been developing the CS and AR variants of the basic model, and for further work, it is planned to develop the other two variants of Fiske's theoret ical proposal and to continue testing the four model variants of pure relational models against the basic one. Some further modifications are planned to advance into the in tegration of all four relational models, together with an agent decisional algorithm to choose what model to use into a particular interaction situation, and the corresponding experiments will also be performed.

We aim to validate and improve this methodology so that the study can be exten ded to other theoretical candidates to be a necessary part of any social simulation. This research programme, in the middle term, could help a wider social community to face up with the Social Scientists Dilemma, and could promote wider access to ABSS methodology.