Novelty in goal-oriented machines using a thread memory structure

Abstract

Resourcefulness and creativity are desirable properties for an intelligent machine. The incredible adeptness of the human mind at seeing situations from diverse viewpoints allows it to conjure many techniques to accomplish the same goal, and hence recover elegantly when one method fails. In the context of goal-oriented machines, this thesis presents a system that finds substitutes for the typical physical resource used to accomplish a goal, by finding novel uses for other, available resources-uses that these resources were not originally meant or designed for. In a domain where an object can serve multiple functions, this requires: (1) understanding the functional context the object is operating in; (2) building a realistic representation of the given objects, which often do not fall neatly into tightly-structured categorizations, but instead share properties with other 'boundary' objects. The system does this by learning from examples, and using the average member, or 'stereotype' as the class representative; (3) allowing imperfection: identifying properties that are not crucial for goal satisfaction, and selectively ignoring them; and (4) measuring similarity between objects to find the best substitute. The system bootstraps with knowledge about the properties of the objects and is given positive and negative examples for the goal. It can infer, for example, that two objects such as an orange (the typical resource) and a ball (the positive example) are related in the context of finding a throwable object on account of their similarity in shape and size, but unrelated in the context of finding an ingredient for a fruit salad, because one is a fruit and the other is not.

(cont.) It then finds a substitute that shares shape and size features with the orange. If, on the other hand, we need an ingredient for a fruit salad, we can supply it another edible fruit as a positive example. The system is implemented in Java; its performance is illustrated with 7 examples in the domain of everyday objects.