A multi-agent systems approach to distributed bayesian information fusion

Gregor Pavlin, Patrick de Oude, Marinus Maris, Jan Nunnink, Thomas Hood
2010 Information Fusion  
This paper introduces design principles for modular Bayesian fusion systems which can (i) cope with large quantities of heterogeneous information and (ii) can adapt to changing constellations of information sources on the fly. The presented approach exploits the locality of relations in causal probabilistic processes, which facilitates decentralized modeling and information fusion. Observed events resulting from stochastic causal processes can be modeled with the help of Bayesian networks,
more » ... ct and mathematically rigorous probabilistic models. With the help of the theory of Bayesian networks and factor graphs we derive design and organization rules for modular fusion systems which implement exact belief propagation without centralized configuration and fusion control. These rules are applied in distributed perception networks (DPN), a multi agent systems approach to distributed Bayesian information fusion. While each DPN agent has limited fusion capabilities, multiple DPN agents can autonomously collaborate to form complex modular fusion systems. Such self-organizing systems of agents can adapt to the available information sources at runtime and can infer critical hidden events through interpretation of complex patterns consisting of many heterogeneous observations. • Constellations of information sources are often not known prior to the operation and they change at runtime. On the other hand the models must capture all observations; i.e., each particular fusion process requires a specific domain model which maps observations from the current constellation of information sources to the hypotheses of interest. Consequently, a domain model should be adapted at runtime, as the information sources become available. • Reliable detection in such settings requires processing of large quantities of noisy information. Given these challenges, a modular approach to modeling and inference seems to be a good choice; adequate domain models can be assembled from basic modeling blocks at runtime and the processing load can be distributed throughout a system of networked devices. It turns out that probabilistic causal models facilitate the design of robust and flexible modular fusion systems. Observations can be often viewed as outcomes of causal stochastic processes. Such processes can be modeled with the help of causal Bayesian networks (BN) [25] which provide a theoretically rigorous and compact mapping between hidden events of interest and observable events. By considering the locality of causal relations in BNs and their factorization properties [13], we derive design and organization rules which support creation of multi-agent systems implementing exact belief propagation in distributed fusion systems. The resulting fusion systems do not require any centralized configuration and fusion control. In addition, no secondary fusions structures spanning multiple agents have to be compiled, which allows flexible configuration at runtime; i.e. the resulting systems support hot swapping and plugging of fusion modules. This is achieved through targeted instantiations of variables in combination with local models constructed according to simple design rules. In particular, we exploit Markov boundaries 1 [20] to systematically find efficient instantiation patterns. The modularization and combination principles are used in Distributed Perception Networks (DPN), a multi-agent fusion architecture. DPN agents are basic building blocks which autonomously form distributed domain models and support decentralized Bayesian fusion through cooperation. Each agent performs a local fusion task and shares its fusion results with other agents. A DPN fusion agent can receive information from various information sources (e.g., sensors, humans) or other fusion agents and computes probability distributions over relevant hypotheses (i.e., belief) by using its local causal models. Outcomes of such local inference processes can in turn be supplied to other fusion agents as input. Depending on the available information sources, DPN agents form a multi-agent system which corresponds to the required causal model. DPN agents thus form a task-specific DPN, which is basically a self-configurable distributed classifier. In terms of [15] , DPN agents implement primarily deductive inferencing based on local fusion algorithms which can be applied at different levels of the JDL model. The presented approach to distributed fusion is complementary to well known approaches to inference with distributed graphical models [29, 17] . In contrast to our method, these approaches require compilation of secondary inference structures, often recursive processes which can be computationally expensive and time consuming. In addition, approach from [17] requires prior knowledge of all information sources. Consequently, these approaches do not support quick adaptation of fusion systems and cannot efficiently cope with domains where information source constellations can change at runtime. The paper is organized as follows: in section 2 we explain why causal probabilistic models are suitable for certain real-world information fusion problems. We also review important properties of Bayesian networks, such as Markov boundaries, which are central to our approach. In section 3 we introduce basic design principles for distributed fusion systems, which are based on instantiation of variables in Markov boundaries. By using the theory of factor graphs we show that agents designed according to these principles can form fusion organizations which support globally correct collaborative information fusion, without centralized configuration and fusion control. In section 4 we introduce Distributed Perception Networks, a multi-agent fusion architecture. Section 5 discusses real world implementation issues, such as robustness with respect to the modeling parameters and assumptions about conditional independences. We also address challenges regarding the generation of complex domain models.
doi:10.1016/j.inffus.2009.09.007 fatcat:xl4hd6k5bbhenl6ro46kr6hkju