Probabilistic reasoning for complex systems

Reasoning under uncertainty is a central issue in artificial intelligence. Real-world agents must deal with noisy sensor information, non-deterministic effects of actions, and unpredictable exogenous events. Probabilistic reasoning methods, and Bayesian networks (BNs) in particular, have emerged as an effective and principled method for reasoning under uncertainty. BNs exploit conditional independence relationships to create natural and compact domain models, thereby supporting useful reasoning patterns, and providing effective probabilistic inference and learning algorithms. However, BNs are inherently limited by their attribute-based nature, making it difficult to apply them to large, complex domains.;This thesis addresses the issue of representing and reasoning about probabilistic models of complex systems. We believe that the key to reasoning effectively about complex systems is to provide a language that supports the expression of system structure. We present a powerful object-based representation language, that integrates logical and probabilistic representations. Our language provides the ability to create structured, modular probabilistic models. The language maintains the key advantages of BNs, exploiting conditional independence relationships. In addition, it is capable of representing other aspects of system structure not represented in BNs. In particular, it supports the decomposition of complex systems into weakly interacting subsystems, and the reuse of models for many different components of a system.;Another key benefit of our language is that it is very flexible. The same probabilistic representations can be applied in many different situations, with very different configurations. In fact, our language can even represent uncertainty over the system configuration itself, and integrate that uncertainty directly with uncertainty over the basic properties of objects in the system. Our framework also supports the representation of powerful recursive probability models.;We present inference algorithms for our language that exploit the structure that can be expressed in it---not only the conditional independence structure normally exploited by BN algorithms, but also encapsulation, reuse of computation and symmetry resulting from the object-based representation. We describe an implemented system that supports representation and reasoning with models in our language, and provide experimental results demonstrating the advantages of exploiting structure in inference.