I will present a series of descriptive-learning algorithms that belong to an alternative, data-driven search paradigm, in which the search is guided not by relationships between the forms of the hypothesized rules but by correlations in the data they represent. Our systems work by exploiting anomalies in this data: They hypothesize that patterns that are very unlikely to have arisen by chance represent features of the domain.
I will describe separate data-driven methods for rule-discovery in propositional and in relational domains. I will also present refinements to our relational rule-discovery system that make possible, first, the automatic identification and exploitation of variable-type information, and second, the use of sampling to increase efficiency at little cost in accuracy.
I will describe how we have applied our methods to the problem of finding planning invariants---formulae that are true in every reachable state of a planning world---which require richly expressive languages to describe. Previous automatic invariant-discovery methods have worked by deduction from the operators. Our discovery methods provide a novel inductive approach to this problem, and can find invariants from merely a few reachable-state descriptions. I will demonstrate that the number and types of laws we discover are comparable to those discovered by systems that require operator descriptions in addition to state descriptions.