Analogy by mapping spreading and abstraction in large multifunctional knowledge bases
Analogical reasoning is one of the most fascinating activities in human thought and can be described as the process of understanding a given situation, called the target, by comparison with another situation, called the source (or base), which is more familiar or better understood. The use of the analogy provides one mechanism for reasoning when classical deductive methods are insufficient. One of the main shortcomings of previous work on analogy is that it required a task-specific knowledge representation, since it can access analogues only if they are already structured in memory, i.e. only if the memory consists of a collection of separately encoded, explicit items. The "standard" analogy algorithm of most previous work is based on a first stage during which a set of potential sources is selected from the knowledge base. If, however, potential sources are not directly identifiable in memory, then this strategy is not applicable. The main objective of this thesis work has been to devise an analogy which can be applied to multifunctional knowledge bases (i.e. "task-general" knowledge bases), which have not been pre-processed for the analogy task. The model of analogical reasoning proposed in this thesis is based on the novel view of the analogue retrieval problem being an analogue construction problem, as opposed to the common view of analogue retrieval as an analogue selection problem. This novel approach combines the access, mapping and generalisation stages of classical analogical reasoning in a single constrained search process, and uses a collection of abstraction operators to achieve a greater mapping flexibility. The model has been implemented in the ALBION system and has been applied to a sizeable multifunctional knowledge base developed independently of this work. This shows the suitability of the model to tackle multifunctional knowledge bases. Additionally, experimental time and space complexity studies indicate that the approach can be efficiently applied to large-scale knowledge bases.