Policy and Complex Systems
own end . This is precisely what happens in ABMs , it follows that we can generate easily and sensibly networks through ABMs .
In addition to the above general consideration , the researches on the topic emphasize a series of limits of NA that could be overcome , thanks to the crossfertilization with ABMs .
The first issue is that of dynamics . De Caux et al . ( 2014 , p . 2 ) point out that much of network theory focuses on static networks , 1 whereas it is obvious that interaction is dynamic and evolutionary .
The second issue concerns the behavior of nodes . NA has to reconcile two different and sometimes apparently irreconcilable aspects : the need of generating a network through appropriate form / severe rules and the need of embedding in such rules a stylized version of meaningful social and economic behaviors . It seems rules that govern the formation of links that the literature in traditional network theory to date employs are usually very straightforward and often lack empirical foundations ( Roth , 2007 ). It follows that , through these models , we can only generate theoretical networks are essentially abstract in nature .
As far as methods are concerned , the traditional mathematical modeling of networks encounters a series of problems . Firstly , the scope for actual interaction is very limited since the behavior of the nodes is synthetized in few formal propositions and this is inherited from cellular automata ; secondly , because of this limitation , the obvious way to explore the possible set of nodes configurations is by means of combinatorics . This leads to a serious problem in mastering the model , since it has been shown ( Johnson & Gilles , 2000 ) that , for instance , a network with eight nodes can generate up to 250 million different theoretical networks .
Considering the dimension of realworld networks , this seems a serious flaw in the possibility of using such models to guide policy decisions . A further consideration is that the use of combinatorics , while mapping all the possible networks , gives no insight about which is more likely to emerge .
To sum up , the process that guides such research is of the following kind : ( i ) take data from real world ( e . g . social media ); ( ii ) observe regularities ( i . e . social networks are often of the small world type ); ( iii ) generate theoretical networks with desired properties ( e . g . stable and efficient networks ); ( iv ) measure the distance between theoretical and actual networks by means of network statistics .
Step ( iv ) is of utter importance . Edmonds and Chattoe ( 2005 ) stress the weakness of the causal association between measures and the actual properties of the whole network in the name of algorithmic non-compressibility :
“ the most individualistic measures ( like density ) are most likely not to capture the overall ‘ flavour ’ of the networks but even for obviously structural measures like centrality and cliques , we are still entitled to ask how well these ‘ subnetwork ’ measures should be expected to capture properties of the whole network ( 2005 , p . 1 ).”
If this is the case , the measures used to perform step ( iv ) might be inaccurate to give an understanding of the social facts that lie behind the network despite the fact that such an understanding that is the ultimate goal of the entire undertaking .
The issues listed so far often show up jointly . For instance , measures of networks can be unreliable due to the inherent dynamic nature of networks . The usual dynamic version of NA consists in generating
79