Bayesian Methodologies Seminar: Graphical Models - From Single Agent to Multiagent

Probabilistic and decision-theoretic graphical models, such as Bayesian networks, provide compact yet powerful formalisms for building intelligent systems, known as agents in the field of artificial intelligence, that must function in partially observable and stochastic environments. After about three decades of active research, a large body of literature is now available on how to construct intelligent agents based on these graphical models. In most of these systems, an agent maintains a graphical model, perceives its environment, uses the model to update its belief and determines its action. Although this single-agent paradigm is sufficient for many applications, there are other applications where it is undesirable or difficult or impossible to assemble the necessary knowledge into a single graphical model, and to grant its access to a single agent for inference. A multiagent paradigm where multiple agents, each equipped with a unique graphical model, cooperate through limited communication, can be better suited to these applications. The talk will outline several advancements in multiagent graphical models, including multiagent equipment monitoring, multiagent collaborative design, multiagent expedition, and multiagent Bayesian forecasting of time series.