Active bias adjustment for incremental, supervised concept learning

Supervised concept learning may be depicted as a search through a space of hypotheses to find (learn) the desired target concept based on a given set of examples of that concept. Typically, each hypothesis covers (implies) all examples of the concept that have been seen. If concept learning is incremental, then hypotheses about the identity of the target concept are formed and then modified with each new example. Due to the large number of potential hypotheses, biases, or hypothesis preferences, are needed to reduce the search space. A bias is good if it allows the concept to be learned easily. Finding a good bias for learning a particular target concept is challenging because the concept is not known beforehand. One solution is dynamic bias adjustment. Although some previous systems have the capability to adjust their bias dynamically, they are limited in their ability to identify erroneous assumptions about the bias. Often, hypothesis language biases are present because of implicit assumptions made by system implementors about the relationship between the bias and the target concept. These assumptions may not always be valid. Without proper diagnosis, it is difficult to identify and then remedy erroneous assumptions. This dissertation introduces the idea of adjusting the bias based on tests of explicit biasing assumptions. Biasing assumption tests use actively requested examples (which are also used for concept learning) for selecting appropriate bias adjustments. These tests provide early recognition of potential learning problems due to a mismatch between the bias and the target concept. Unlike previous work, this work has formalized a basic set of biasing assumptions. The formalization has provided a foundation for proving useful theorems and for implementing a system that tests and dynamically adjusts its bias. The implemented system has provided an empirical demonstration that adding a combined approach of active biasing assumption testing and dynamic bias adjustment to a concept learner can lead to a three-fold improvement in the rate of convergence to the target concept.