Building successful Artificial Intelligence (AI) systems for real-world domains might only be possible by incorporating a flexibility that is realized through learning. As a result, machine learning studies in real-world environments are gaining prominence (Carbonell & Hood, 1985; Langley, Kibler, & Granger, 1985; Sammut & Hume, 1985). These studies relax many of the simplifying assumptions that enabled earlier work to progress. In particular, this proposal is concerned with incremental concept learning systems, which drop the commonly-made assumption that all environmental observations can be processed simultaneously (Dietterich & Michalski, 1983; Hayes-Roth & McDermott, 1978; Quinlan, 1986; Vere, 1980). Instead, observations are assimilated as they become available. This incremental assumption has important computational ramifications, for it bounds the amount of processing that can be expended on new observations.
Figure 1: A Model of Intelligent Processing
A simple model of intelligent processing is depicted in Figure 1 (Dietterich, 1982). Learning organizes environmental observations to improve performance on some task(s). Assumptions about the environment, knowledge base, and performance task impact the formulation of the learning process, but it is the environment that primarily plays the role of the `independent variable.' The environmental assumptions that underlie nonincremental learning have crept into all aspects of the intelligent processing model. In particular, nonincremental systems assume that all `required' information is available from the outset and conceptual knowledge is induced `all-at-once'. Since nonincremental systems do not assume that knowledge is intermittently updated and applied, the learning element may extensively search for appropriate concepts with little concern for efficiency. Moreover, induced concepts need not be organized for efficient retrieval - the result is a simplified `knowledge base' that is little more than a list of disparate (often a single) concept descriptions. Finally, the extensive search invariably carried out by nonincremental systems insures that performance (e.g., prediction accuracy) tends toward optimality.
While nonincremental systems may lead to near-optimal performance, their environmental assumptions and processing characteristics become increasingly untenable as the number and variety of observations grow. We will study the cost/performance tradeoffs implied by incremental learning. This study will focus on a number of behavioral, computational, and representational issues.
In summary, our primary goal is to investigate domain-independent mechanisms and representations that efficiently and robustly support incremental learning. However, the applicability of our investigations promises to reach beyond the context of incremental learning. For example, guidelines for identifying and maintaining default generalizations during learning may have important implications for belief revision. Further, recent studies demonstrate that induced knowledge may `over-fit' the data and actually detract from performance. We will explore the apparent tradeoff between knowledge `complexity' and performance ability and its implications for incremental learning.
Second, we will test our general methods on real-world tasks such as intelligent tutoring where intermittent, accurate, and timely feedback is invaluable.
Finally, we have and will continue to empirically evaluate our learning systems by methods addressed under the research plan. Importantly, our methodology downplays anecdotal evidence and insists that AI systems be characterized across a range of possible scenarios. Apparently, our approach has already had some impact on the methodological biases of the machine learning community (Langley, 1987b).