Section: New Results

Peer-to-Peer Inference Systems

Participants : Nada Abdallah, Vincent Armant, François Calvier, Philippe Chatalic, Philippe Dague, François Goasdoué, Chantal Reynaud, Laurent Simon.

A P2P inference system (P2PIS) is made of autonomous agents called peers. Each peer models its application domain using a knowledge base (KB) and peers having similar interests can establish semantic correspondences between their KBs called mappings. Mappings play a central role since on the one hand they define how the KBs of some peers integrate and on the other hand they give rise to a semantic network, the decentralized KB of the P2PIS, in which it becomes possible to reason. However, reasoning in the distributed logical setting of a P2PIS is not simple. Indeed, the challenge is to design decentralized reasoning algorithms with the purpose to reduce an inference task to perform on the KB of a P2PIS to a decentralized calculus among the peers, while none of them has a comprehensive view of the global KB.

Consequence Finding

In the last years, we have investigated the basic AI task of consequence finding in (possibly inconsistent) propositional P2PISs. Consequence finding consists of deriving theorems of interest that are intentionally characterized within a logical theory. It proves useful in many composite AI tasks such as common sense reasoning, diagnosis, or knowledge compilation. Most of our results have been implemented in the SomeWhere platform, the scalability of which has been demonstrated on synthetic data (up to a thousand of peers). It is worth noticing that our results allow one for the very first time to perform a truly decentralized consequence finding calculus in a distributed theory of propositional logic, i.e., without having a global view of that theory.

SAT Solving

Following our previous work, we identified a way of predicting learnt clause usefulness in modern SAT solvers. This work [16] allowed us to propose a new version of Minisat that won the industrial UNSAT category in the 2009 SAT Competition. This contest category is highly competitive.

P2P conservative extension checking

We have pointed out that the notion of non conservative extension of a knowledge base (KB) is important to the distributed logical setting of propositional P2PIS. It is useful to a peer in order to detect/prevent that a P2PIS corrupts (part of) its knowledge or to learn more about its own application domain from the P2PIS [5] . That notion is all the more important since it has connections with the privacy of a peer within a P2PIS and with the quality of service provided by a P2PIS. We have therefore studied the following tightly related problems from both the theoretical and decentralized algorithmic perspectives: (i) deciding whether a P2PIS is a conservative extension of a given peer and (ii) computing the witnesses to the corruption of a given peer's KB within a P2PIS so that we can forbid it.

P2P consistency checking and query answering

We have investigated a decentralized data model and associated algorithms for peer data management systems (PDMSs) based on the DL-lite description logic [9] , [10] . That logic is a fragment of the forthcoming W3C recommendation for the Semantic Web: OWL2. Our approach relies on reducing query reformulation and consistency checking for DL-lite into reasoning in propositional logic. This enables a straightforward deployment of DL-lite PDMSs on top of SomeWhere, our scalable peer-to-peer inference system for the propositional logic. We have also shown how to answer queries using views – predefined queries – in DL-lite in the centralized and decentralized cases, by combining the query reformulation algorithm of DL-lite and the state-of-the-art query rewriting algorithm: MiniCon.

Tools for experimental evaluation of P2PIS architectures

Because of the distributed nature of SomeWhere like P2PIS and their assynchronous mode of communication, performing large scale experiments with such architectures is a complex tasks. In order to alleviate this task, we are developping a set of tools developed in order to the automated deployment of such P2PIS on the Inria Grid'5000 plateform. SWTools contains a generator of P2PIS instances, with random (but convincing) local theories and mappings. It allows for multi-cluster node reservation on the grid, the automatic deployement of peers on these nodes, query generation and dispatching on the network and automatic results collection on the peers.

Distributed Diagnosis

Research on consistency-based distributed diagnosis, set up in the framework of propositional P2PISs, pursued in 2009. It is in some sense a dual problem of consequence finding, as the algorithm developed is based on the distributed computation of prime implicants of the (unknown) global theory (so, not relying on a preliminary computation of conflicts as most of the diagnosis algorithms in the centralized case). We improved our first algorithm, which incrementally returns diagnoses by dynamically building a tree throughout the network. by taking advantage of a jointree structure. We have implemented an automated benchmark generator which builts peer-to-peer inference systems structured by social network topologies. Experimentations and comparison of the approaches are an on going work, as well as addressing scalability issues. We handle also privacy respect of the peers by allowing agents to reason “as much as possible” together while keeping their secret secret. Some privacy of the final diagnostic results is also studied. Up to now, the network is considered as static, i.e. the acquaintances of each peer are fixed. Next step will be to study dynamicity of the network, i.e. addressing departures and arrivals of peers.

This research is one of the topics of the submitted proposal of associated INRIA team Smarties with INRIA Rennes Dream group and NICTA Canberra (Australia).

Mapping distributed ontologies

Our work takes place in the setting of the peer data management system (PDMS) SomeRDFS. Ontologies are the description of peers data. Peers in SomeRDFS interconnect through mappings which are semantic correspondences between their own ontologies. Thanks to its mapping a peer my interact with the others in order to answer a query. A crucial aspect in SomeRDFS is that peers are equivalent in functionalities. No peer has a global view of the data management system. Each peer has its own ontology, its own mappings and its own data. It ignores the ontology, the mappings and the data of the other peers. In this setting, our work aims at increasing the mappings of the peers in order to increase the quantity and the quality of the answers of the whole data management system. Previously, we proposed an approach to identify two kinds of mappings: mapping shortcuts corresponding to a composition of pre-existent mappings and mappings which can not be inferred from the network but yet relevant. We focused on the identification of the second kind of mappings. We updated our algorithms identifying relevant mapping candidates and proposed new filtering criteria to limit the matching process to a restricted set of elements. These criteria are relative to the peers involved in the mappings, the kind of mappings which are looked for (generalization or specialization mappings)or to the quality of the mappings. Then we focused on the alignment step. We proposed alignment techniques suitable to our context. We adapted terminological alignment techniques, proposed to integrate mechanisms based on query answering and techniques using the PDMS as an external source. Finally we investigated a methodology to help validating discovered mappings. These new propositions have been implemented in SpyWhere and experiments are currently conducted. This work is the core of a PhD which will be defended next year.