Reviewed by
Paolo Turrini
University of Siena and LABSS, Laboratory on Agent Based Social Simulation, ISTC, CNR Rome
Why do organizations need regulations to adapt and survive? And why is it not enough to issue norms in order to achieve regulation? Why do power relations and roles need to be understood in order to study organizational life? What is to be socially expected after a convention is spread? C.O.I.N. in M.A.S. book provides many answers, with interesting explicit and implicit confrontations between them and two clear core ingredients: socio-cognitive realism and formal method.
Not many years ago the mainstream interpretation of norms in computer science (Shoham and Tennenholtz 1995) viewed regulation as mere physical constraint on agents' behaviour. The approach of "obedience by design" has been recently criticized both for its cognitive unrealism (Conte and Castelfranchi 2006) and its computational costs (Governatori et al. 2006). Contributions in this book show how how further computer science can get much further in capturing and rigorously describing interaction dynamics within and between organizations.
On the one hand the importance of interaction constraints by means of objects like norms or conventions is argued for. Most contributions claim that agents need these objects in order to pursue their goals in a more efficient way. Among the many possible ways to look at the problem, many authors choose the normative approach, based on the use of norms in artificial institutions. The meaning of the notion of artificial institution is discussed and clarified throughout the book.
On the other hand, several contributions focus on the dynamics of adaptation and coordination in heterogeneous organizations by means of formal specifications for organization-centred programming, paying particular attention to the notion of agents' autonomy and to its relation to the need for social order.
The papers in this volume are drawn from two complementary events, ANIREM (Agents, Norms and Institutions for Regulated Multiagent Systems) and OOOP (From Organizations to Organization-Oriented Programming in MAS), that were part of the workshop program at AAMAS 2005 in Utrecht.
Even though ANIREM and OOOP have been designed as separated events their commonality of interests, the strong intersection between their Program Committees and the computer science background of their Workshop Organizers, fully explains and justifies the regrouping of contribution within four sections that will be here analyzed: (I) Modelling Analyzing and Programming Organizations; (II) Modelling and Analyzing Institutions; (III) Modelling Normative Designs; (IV) Evaluation and Regulation.
In the first part, Modelling Analyzing and Programming Organizations, the problem of a theory of "sociology of organized action" stated in Sibertin-Blanc et al. (3) is somehow dealt with by all authors, though from different perspectives. In line with agent-based social theory (Coleman 1990; Conte Castelfranchi 1995), organizations are grounded on interdependent agents that interact for common or conflicting goals. The notion of team, idealistically assumed as "cooperative, reliable and having non-conflicting norms", is proposed in order to reason on sustained effectiveness (Wijngaards et al., 42).
Agents organizations are generally not seen as flat groups of agents but ontologies and micro-macro analyses are provided. In particular, according to Van den Broek et al., the notion of role is seen as the basis for understanding organizational dynamics like information flows and environment adaptation. Organizations do not differ only in the micro-macro levels but they change through time. And McCallum et al. provide a formal language for the analysis and the verification of this change. In particular artificial and natural societies are constantly compared and the notion of agent autonomy comes as expected into play, especially in Hubner et al..
All contributions are somehow far from a simplistic obedience by design, and all pose the problem of how an organization needs to interact with its lower elements, providing basis for developing a rigorous understanding of the notion of autonomy in organizational settings and in general for building formal understanding of many still not formally investigated notions. In this respect, Sibertin-Blanc et al. (16) point out how AI and MAS need to produce "new theories of mental and social phenomena that can benefit to sociology".
The section Modelling and Analyzing Institutions is opened by Pablo Noriega's invited paper that, on the grounds of the pioneer work of IIIA with electronic institutions (Noriega and Sierra 2002), describes how the regulatory objects, that he calls conventions, have to be not only established, but also enforced and maintained.
Moreover, after an analysis of electronic institution cases, he proposes to evaluate organization against four dimensions: grounding, the relation between actions taking place within an institution and the real world; degree of agentification, that is the need for software agents within an institution, as constitutive components or as participants; autonomy, that in electronic institutions is translated as the extent of requirements imposed on the agents and the capacity of an organization to enforce regulated interaction.
His intuitions provide indeed "a set of distinctions that can be beneficial for the development of applied regulated agent systems" (97).
The papers that follow deal with the analysis and the specifications of agent-based institutions. In particular (Cliffe et. al) propose a non-monotonic logic programming language for modelling reasoning and verification tasks. Agent based institutions are here defined with respect to institutional facts, distinguished further in institutional domain facts (institution-dependent) and normative facts, such as institutional power, permission, obligation, violation. The stress on executable specifications and the formal notions of prediction, post diction and planning provide useful and solid categories for checking possible institutional states. The need for expressing sanctions and recovery from violation is added to the authors' wish list.
Much of these issues are addressed Boella et al, where the notion of autonomy and the possibility for normative agent to deviate from the ideal are emphasized. Autonomous mental states are taken into account and norms are related to them. Also Rubino et al., by introducing introduce the notion of computational institution, grounded on Noriega and Sierra electronic institution concept, provide tools for verifying compliance to norms. In Vigano' et al. instead conventions represent the concrete performance of institutional actions while norms impose obligations and restrictions for the agents that interact within the institutions.
Modelling Normative Design is indeed the most operational part, with innovative contributions aiming at filling the gap between deontic description and operational protocols (a gap which is indeed extremely important to fill), and addressing notions like norm internalization and the way organization affects agents' behaviour. In particular, in Aldewereld et al. (a, 157), it is shown how protocols provide "a way to reduce complexity by giving direct, step by step guidelines for behaviour, as long as the protocols comply with the norms", which represents indeed an advance for MAS regulation. Their tools are yet applied to Electronic Institutions, and the paper contains an explicit statement of operational inadequacy of norms as expressed in deontic logic, for they say only what is acceptable and not how to achieve it. Aldewereld and colleagues tell us how to achieve it. The temporal dimension embedded (they use Computational Tree Logic) in the formal language makes it intuitive and expressible, especially in its relation with the notion of deadline. Norm obedience and goal directedness are translated into the classical CTL notions of safety and liveness that make this this notions easily checkable (see Clarke et al. 1999).
The both technical and philosophical investigation of norms carried out by all the authors goes on with (Garion and van der Torre) that, by formally implementing design by contract between callers and routines, show how "violations and exceptions are distinct concepts that should not be confused" (170). They moreover provide, in terms of contract violations, one further distinction between strong acceptability level and weak acceptability level concerning the caller satisfaction level.
This notion is further analysed even though from different perspectives by Kollingbaum and Norman, who deal with deliberation in norm governed practical reasoning. Very interesting questions are raised by the authors, for instance, "can the normative authority, which issued such norms, be convinced to revoke existing prohibitions or obligations or at least temporarily grant a permission that overrides a prohibition?" whose answer would represent a strong advance in understanding normative change.
Finally, Boella and van der Torre implement a social law as a set of restriction on the agents' activities. Even if this concept is very simplistic, they empower it with the use of counts as conditionals and semantics of organized games, reducing the understanding on norm effectiveness to game theoretical interaction (see Coleman 1990).
The last section, Evaluation and Regulation, is the shortest one but it yet contains a large amount of insightful advances that range from a simulation framework of organizational change (Dignum et al.) to a formal approach for the verification of norm compliance (Aldewereld et al., b) and to a temporal hybrid language for defining social expectations based on metric interval temporal logic (Cranefield).
In Dignum et al. simulation provides insight concerning whether a reorganization should or should not take place, and evaluating the congruence between task performance and organization. Agents are once again taken as autonomous entities, which are able to determine their action on the grounds of their own internal states.
Authors show how "greater cognitive realism in social simulations may make significant differences in terms of organizational performance" (214). They distinguish and reason about behavioural (short term) change and structural (long term) change and identify those factors that determine the need for reorganization. The employed methodology is also very interesting, for a big effort is put in order to have the highest level of cognitive realism, together with a formal specification before the implementation, that render the simulation results understandable and easy verifiable.
Aldewereld et al. (b) build a formal method for verifying the norm compliance of protocols. Using Linear Time Temporal Logic authors keep the temporal framework of protocols specifications and formalize norms and related notions. In particular the notion of permission, which is seen as exception to a general prohibition, is dealt with in a strongly innovative way. The formalism is used to represent practice and to provide useful tools for norm compliance verification.
Concluding the book, Cranefield yet remains outside the "compliance by design" paradigm, and provides an extremely powerful and innovative hybrid temporal language (see Blackburn et al. 2001) to verify norm compliance of agents societies, together with algorithms for process monitoring.
With a wide range of topics and 17 selected papers, the book poses many questions and provides interesting innovative answers that are of use for computer scientists, but also for economists, policy makers, philosophers, sociologists and cognitive scientists.
The book preface writers claim that "the result is a well-balanced collection of high-quality papers that really can be called representative of the field at the moment". I do agree with them. Nevertheless, there still seem to be much disagreement on which are the very key factors needed to explain regulation, their names and mechanisms, what is the ontological status of regulatory objects like conventions and norms, how do they work, and which of the many perspectives on agent interaction is the fundamental one to understand organizational regulation.
Disagreement is a necessary factor for scientific debate, and thus for scientific growth, that in the field need to be translated into the ambition of building a unified theory of Coordination, Organization, Institutions and Norms. The way it can be done is stated by Oskar Morgenstern's seminal work on organizational theory (Morgenstern 1951):
"desirable theory would undoubtedly have to be quantitative in character and necessarily highly mathematical, and comprise the great variety of forms in which the object appears in reality" (5).
Indeed, socio-cognitive realism and formal method have been working tools for all the authors of this book, and they allowed producing scientific statements, easily checking their validity, and building up precise notions that help discussion and interdisciplinary understanding.