Belief revision and machine discovery
Donald Rose, Pat Langley · 1989
In order to account for new observations, an intelligent agent must change those beliefs when they have become inconsistent due to these new data. In particular, scientists are often forced to revise theories when the beliefs they are based on conflict with new experimental evidence. For automated discovery systems to exhibit this ability, they must also have mechanisms for revising their beliefs in order to regain a globally consistent database. To this end, we have developed a framework for integrating the processes of discovery and revision, and instantiated the framework in a computer program named REVOLVER. This system employs general rules for scientific discovery and belief revision, and works in domains where knowledge can be codified as reactions among objects. When inconsistencies arise, the program employs a form of heuristic search known as hill climbing in order to find a new consistent theory (set of beliefs). In this thesis, we first discuss the motivation for our research, along with the goals and tasks addressed by REVOLVER. Next, we discuss previous work in discovery and revision which directly influenced our design of the system. We then describe the program itself, illustrating its representation, its basic rules and inference process, and its method for belief revision. Next, we show the system's generality through its replication of discoveries in historical domains. We then provide a more detailed analysis, discussing experiments that illustrate REVOLVER's robustness when faced with various artificial domains. After discussing related work, we present a concluding chapter that discusses the contributions and limitations of this work, ideas for future enhancements, and final thoughts.