Supporting conceptual knowledge capture through automatic modeling

One of the main goals of science is to gain better understanding of how nature works through observations and experimentation. As such, science does not concern itself so much with facts, but with unifying theories that explain why facts are the way they are. Models can be seen as (partial) implementations of theories and are means towards proving the correctness of a theory. Constructing such models automatically is a long-term research goal of the Machine Learning area within Artificial Intelligence. Consider the issue of automatic model building. A program solving this task would take a set of facts as input, generate a model based on these facts, and use this model to make predictions. As such, machine learning can be considered an automatic model building approach. However, most machine learning algorithms focus on the accuracy of their predictions, and less on providing an explanation for their predictions (i.e. a learned model often provides little insight into the causality underlying the system's behaviour). One of the aims of Qualitative Reasoning (QR) is to represent the conceptual knowledge experts have of systems, particularly concerning causality, in a way that explanations about the behaviour of these systems are captured [1]. Garp3 (http://www.garp3.org) is a workbench that can be used by experts for building such models [2,3]. This software was further developed within the NaturNet-Redime project (http://www.naturnet.org/) and is used by domain experts to capture their insights on environmental science (e.g. [4,5,6]). However, building models is in general difficult and time-consuming (and QR models are no exception to this), hence the need to support this process, particularly by automating it as much as possible. This paper discusses our work on an automated modelling algorithm that learns Garp3 models based on a qualitative description of the system's behaviour. In contrast to other approaches, our algorithm attempts to learn the causality in the system in the form of causal dependencies. The algorithm uses consistency rules to determine the causal dependencies that hold within the system. Using the concept of clusters the search space is significantly reduced. The algorithm achieves good results when used to automatically re-create already existing and well-established models. This suggests that it is possible to automatically infer causal explanations from qualitative behaviour descriptions of a system, and that model-building support through an automatic model-building algorithm is viable. In the paper we will describe the algorithm, the results it produces on automatically generating models, and our plans for future work. The latter particularly focuses on an interactive version of the software, so that users can steer the automated modelling process to their own liking.