Primitive concept formation
Kevin B. Korb, Christopher David Thompson · 2002
Our goal is to demonstrate the feasibility of an autonomous learning agent by developing means to learn and employ concepts in a primitive machine intelligence which must operate in a real-time, uncertain (noisy) environment. The paper reports on the first steps towards such an agent: the development of an agent, Alice, who starts out with only a primitive set of concepts-corresponding to perceptible attributes of objects in the environment and to her own utility function-and who generates a conceptual structure using cognitively plausible rules of concept formation and refinement, abstracting from the immediate attributes of the mushrooms she finds and their longer-term impact on her utility, in a goal-driven manner. The concept formation rules we have developed are more conservative than such standard methods of concept formation as version space methods and ID3. We suggest that this caution offers a competitive advantage in difficult environments.>