PhD Thesis
Computer Scienze Institute
University of Ancona
1998

 

Belief Revision in Multi-Agent Systems


 


Thanks to the fast progresses in hardware and communication technology, today it is possible to connect inexpensively a large numbers of powerful processing units that execute asynchronously. This gives us the possibility of design complex systems in which different intelligent unites work together to solve problems that are beyond their individual capabilities. In last 15 years, Distributed Artificial Intelligence has faced this matter with Multi-Agent Systems(MAS). The focus of MAS research is on the agent (getting it to interact meaningfully with other agents) and on collections of agents as a whole (identifying under what conditions the agents will rationally choose to act, such that the society displays certain properties). Agents are conceived as autonomous and rational entities that perform tasks in complex and dynamically changing environments: autonomous because have the ability to make decisions based on aninternal representation of the world, without being controlled by a centralized instance; rational because it is assumed that an agent will act in order to achieve its goals. Autonomy and rationality are attributes strongly dependent on the consistency of the agent's knowledge, and in particular, they can not prescind from it.

By belief revision it is meant the process of rearranging a cognitive state in order to embody an incoming information while preserving the global consistency. Since the seminal, philosophical and influential works of Gärdenfors et al. [83,105], ideas on belief revision have been progressively refined [85] and ameliorated toward normative, effective and quasi-computable paradigms [82,130,131]. Some of the main contributes have been the idea of "revision as transmutation of partial epistemic rankings" [91], "revision for finite bases" [104], the distinction between revision and updating [92] and a related approach based on the notion of possible models [94]. Side by side to this "symbolic" line of research, there has been also a "numerical" way to belief revision [111] whose main contributes were the probabilistic [109], possibilistic [114], and the evidence-based approaches [110]. 
In this dissertation, we propose two ways to perform belief revision in multi-agent environments. In the first case the belief revision problem is approached from a centralized perspective, that is, it is assigned to a single agent; whereas, in the second case, it is approached from a distributed perspective, that is, it is assigned to a group of agents. 
According to us, in a multi-agent environment, where information come from a variety of sources with different degrees of reliability, belief revision has to depart considerably from its original framework. Particularly, the principle of "priority of the incoming information" should be abandoned; while it is acceptable when updating the representation of an evolving world, that principle is not generally justified when revising the representation of a static situation. In this case, the chronological sequence of the informative acts has nothing to do with their credibility or importance. Another point is that changes should not be irrevocable. To make practical and useful belief revision in a multi-agent environment, we substitute the priority of the incoming information with the principle of recoverability: any previously believed information item must belong to the current cognitive state if it is consistent with it. 
Our approach to centralized belief revision combines symbolic and numerical operations. Particularly, we use operations in Assumption Truth Maintenance System [122] style to treat the symbolic part of information, and the Dempster-Shafer theory [116] to treat their numerical part. We distinguish two knowledge bases: the knowledge background (KB), which is the set of all the pieces of knowledge available to the reasoning agent (since it can be inconsistent, it cannot be used as a whole to support reasoning and decision processes); and, the knowledge base BÍ KB, which is the maximally consistent, currently preferred piece of knowledge that should be used for reasoning and decision supporting (since it is maximally consistent, it can contain incredible pieces of knowledge). The incoming information, with its weight of evidence, is confronted not just with B, but with the overall KB. 

Belief revision is a process computationally very expensive. From an engineering point of view, this means that it not always convenient or possible to build systems in which the belief revision process is performed only by a central unit. Therefore, an alternative approach is the distributed one, where the belief revision process is no longer assigned to a single agent but to a society of agents, in which each single is able to perform belief revision and communicate with the others. The main difference is that in this case there is no more a single agent having a global view of the system, but any agent has a partial view of it. This allows that the computational load to be divided among the agents of the group. 
Our way to distributed belief revision consists of a society of agents, which adopt locally a centralized belief revision model, communicate with each other and adopt global strategies to extract an emergent global opinion regarding the treated information. In particular, we assume that the agents locally adopt our model for centralized belief revision, spontaneous and/or on-demand polities of communication and data or agent driven election to form global opinions. This approach is conceptually different from the classical Distributed (Assumption) Truth Maintain Systems [124,125] since it allows each agent to maintain its own opinions, independently from those of the others. We determinate two main desiderata: convergence (i.e., which local decision strategies favour the convergence of the opinions rather than the chaotic divergence or the static indifference?) and correctness (i.e., which local decision strategies favour the convergence of the opinions to the most correct pieces of knowledge globally available to the overall agency?). Obviously, the choice of the best strategies depends on the structure of the agency and on the context to which it is applied. 
We believe that the evaluation and the comparison of the performances of the centralized and distributed approach can be done only on a simulative basis. At the end of this dissertation, we will present a study on the simulative basis of the behavior of a society of agents that interact in presence of conflicting knowledge. Particularly, we will try to evaluate and compare the performances of the two approaches to belief revision. Furthermore, we will study how the cognitive state of each single agent evolves during interaction, and how these single evolutions influence the global knowledge of the group. 

This dissertation is structured as follows:

Chapter 1 illustrates and compares the two main research areas of Distributed Artificial Intelligence: Distributed Problem Solving and Multi-Agent Systems. Particularly, it evidences the main characteristics that differentiate an agent from a problem solver. 
Chapter 2 gives a general perspective of belief revision. It starts by describing and comparing the main theoretical contributes of the symbolic and numerical approaches and finishes by giving a brief description of the two main practical approaches. 
Chapter 3 presents our two models for belief revision. It introduces the principle of recoverability and the reasons that make it very suitable in a multi-agent environments. 
Chapter 4 presents a study on the simulative basis of the behavior of a society of agents that interact in presence of conflicting knowledge. It gives a brief description of our specific belief revision test-bed for multi-agent systems and reports the description and the results of five different types of experiments run with it. Moreover, it outlines general considerations on the performances of the adopted strategies.
 


For an electronic copy of this thesis, please contact me pgiorgini@cs.unitn.it