The Übercruncher: Concept Formation by Analogy Discovery

Concept formation allows for knowledge transfer, generalization, and compact representation. Concepts are useful for the creation of a generally intelligent autonomous agent. If an autonomous agent is experiencing a changing world, then nearly every experience it has will be unique in that it will have at least slight differences from other experiences. Concepts allow an agent to generalize experiences and other data. In some applications, the concepts that an agent uses are explicitly provided by a human programmer. A problem with this approach is that the agent encounters difficulties when it faces situations that the programmer had not anticipated. For this reason, it would be useful for the agent to automatically form concepts in an unsupervised setting. The agent should be able to depend as little as possible on representations tailored by humans, and therefore it should develop its own representations from raw uninterpreted data. I propose to develop an algorithm which will use unsupervised analogy discovery to create a conceptual structure (also known as an ontology) that’s useful for concisely characterizing data. One purpose of concept formation (and abstraction in general) is to concisely characterize a set of data (Wolff 2003). With this view, one can use minimum description length as a guiding principle for concept formation. I propose to use this principle to form an ontology of concepts from a collection of data. The data is a set or a stream of statements, where each statement is an ordered set of symbols. These symbols have no meaning for the program other than they’re considered to be ground statements. For example, these symbols can be raw sensor data, or raw descriptions of chess games. My hypothesis is that this algorithm will serve as a general unsupervised learning system that requires no domain knowledge or human programmed representations aside from encoding the data as a set of ordered sets of symbols. This means that this algorithm should produce useful concepts whether the data is from robot sensors, a description of a board game, or a simulation of automobile traffic. However, since the algorithm will be “bootstrapping” its representations from data, the data should be cheap and plentiful. In particular, there should be sufficient data for any useful patterns to be manifested. Since my method uses the same representation framework for raw data as it does for higher