A more historical-cognitive approach was the aim of the work on BACON, which rediscovered various scientific laws by finding patterns in numerical data (Langley, Simon, Bradshaw & Zytkow, 1987). Simon's early work on finding patterns in sequences (Simon & Kotovsky, 1963) was extended in BACON to heuristic search for patterns in numerical data. The most creative of BACON's abilities was the decomposition of relational data to conjecture intrinsic properties in one or more of the objects engaging in the relations. This step went beyond curve-fitting and was based on the metaphysical assumption that an entity's relational properties are caused by its intrinsic properties. In addition to the data-driven tasks modeled in BACON, the group also investigated theory-driven discovery in STAHL.

Over the past decade, most of my discovery research has focused on a new framework, inductive process modeling, that combines background knowledge in the form of generic processes with time-series data to construct explanatory models stated as sets of differential equations. The basic approach carries out exhaustive search through a space of model structures followed by gradient descent through the parameter space for each candidate structure. Later work extended the framework to use constraints among processes to guide search through the structure space and even to induce constraints to discriminate between successful and unsuccessful structures.