Reference: T. K. S. Kumar. A Biological Motivated Algorithm Using Compositional Modeling To Generate Explanations. Knowledge Systems Laboratory, 2000.
Abstract: Much of the work related to generating explanations has concentrated on new methods for modeling and simulation, such as qualitative physics, in which the device models are specifically crafted for explanatory purposes. There has been recent work on Compositional Modeling which provides a formalism for constructing models from modular pieces. This approach provides a number of advantages over conventional simulation software mainly because of its causal interpretation of data. However, little work was done with regard to a supporting algorithm that can generate cogent explanations from the simulation values and causal graphs. Earlier attempts did not solve the problem of irrelevant details introduced by using compositional modeling, as a result of which misleading references resulted in attempting explanation of device behavior. This was mainly because they were based merely on equation tracing and did not try to infer anything about the working phenomena from the causal order graph. We present an algorithm that infers the active working template phenomena based on ideas of mutation and evolution, independent of the domain. This poses many advantages over just working with the quantities defined in the equation model. A byproduct of this is in capturing the user's psychology in terms of phenomena rather than in terms of mathematical equations defined by some other person. The explanation is in the form of natural language rather than graphs of numerical variables. We also describe a number of extensions of the algorithm to handle issues such as scalability and ranking by significance.
Full paper available here.