Computational Ideation in Scientific Discovery: Interactive Construction, Evaluation and Revision of Conceptual Models

We present several epistemic views of ideation in scientific discovery that we have investigated: conceptual classification, abductive explanation, conceptual modeling, analogical reasoning, and visual reasoning. We then describe an experiment in computational ideation through model construction, evaluation and revision. We describe an interactive tool called MILA–S that enables construction of conceptual models of ecological phenomena, agent-based simulations of the conceptual model, and revision of the conceptual model based on the results of the simulation. The key feature of MILA–S is that it automatically generates the simulations from the conceptual model. We report on a pilot study with 50 middle school science students who used MILA– S to discover causal explanations for an ecological phenomenon. Initial results from the study indicate that use of MILA–S had a significant impact both on the process of model construction and the nature of the constructed models. We posit that MILA–S may enable scientists to construct, evaluate and revise conceptual models of ecological phenomena. Background, Motivations, and Goals We may adopt several epistemic views of ideation in scientific discovery. We have developed computational techniques and tools for supporting some of these epistemic views. Here we briefly present these epistemic views as background and motivation for the present work. Conceptual Classification Classification of data into concepts is ubiquitous in science. We all know about Linneas’ classic work on classification in biology. Classification continues to be important in modern biology (e.g., Golub et al. 1999). Classification also is one of the most studied topics in cognitive science, artificial intelligence and machine learning (e.g., Langley 1996; Mitchell 1997; Stefik 1995; Thagard 2005; Winston 1993). The classic DENDRAL system (Lindsay et al. 1980) classified mass spectroscopy data into chemical molecules that produced the data. Chandrasekaran & Goel (1988) trace the evolution of early knowledge-based theories of classification. We have studied both top-down hierarchical classification in which a concept is incrementally refined based on data (Goel, Soundarajan & Chandrasekaran 1987), and bottomup hierarchical classification in which features of data are incrementally abstracted into a concept (Bylander, Goel & Johnson 1991). In recent work, we have developed a computational technique that grounds the concepts in bottom-up classification in perception and uses metaknowledge of this perceptual grounding for repairing the semantics of the concepts when the classification results in an error (Jones & Goel 2012). The Augur system uses this meta-reasoning technique for revising almost-correct classification hierarchies in a variety of domains. Abductive Explanation Scientific theory formation often entails abductive inference (Magnini 2001), i.e., inference to the best explanation for a set of data. Artificial intelligence research has studied abduction from multiple perspectives (e.g., Charniak & McDermott 1985; Josephson & Josephson 1996). The classic BACON system (Langley et al. 1987) abduced physical laws, such as the gas law, from data. Bylander et al. (1991) have analyzed the computational complexity of the abduction task. We have studied computational techniques for abductive explanation that assemble composite explanations for explaining a set of data from elementary explanations that explain subsets of the data (Goel et al. 1995). The RED system uses this technique for identifying red-cell antibodies in a patient’s blood serum (Fischer et al 1991). RED uses domain-independent heuristics such as the essentialness heuristic for assembling a composite explanation from elementary explanations: this heuristic says that if some data item can be explained by only one elementary hypothesis, then the hypothesis should be directly included in the composite explanation.