Why Experimentation can be better than `Perfect Guidance'

Tobias Scheffer, Russell Greiner and Christian Darken

Proceedings of the Fourteenth International Conference on Machine Learning (IMLC-97), Nashville, July 1997

Many problems correspond to the classical control task of determining the appropriate control action to take, given some (sequence of) observations. One standard approach to learning these control rules, called behavior cloning, involves watching a perfect operator operate a plant, and then trying to emulate its behavior. In the experimental learning approach, by contrast, the learner first guesses an initial operation-to-action policy and tries it out. If this policy performs sub-optimally, the learner can modify it to produce a new policy, and recur. This paper discusses the relative effectiveness of these two approaches, especially in the presence of perceptual aliasing, showing in particular that the experimental learner can often learn more effectively than the cloning one.

Explanation-based learning

Explanation-Based Learning (local copy)

Russell Greiner

MITECS (MIT Encyclopedia on Cognitive Science), 1998.

"Explanation-based learning" (EBL) systems attempt to improve the performance of a PROBLEM SOLVER (PS), by first examining how PS solved previous problems, then modifying PS, to enable it to solve similar problems better (typically, more efficiently), in the future. This article first motivates and explains this explanation-based learning task, then discusses the challenges that any EBL system must address, and presents several successful implementations. It concludes by distinguishing this learning task from other types of learning.

(See also material in EffectivePerformanceSystems.)

Hierarchical Classifier

Learning hierarchical classifications

Dale Schuurmans, Adam Grove, Russell Greiner

Many significant real-world classification tasks involve a large number of categories which are arranged in a hierarchical structure; for example, classifying documents into subject categories under the library of congress scheme, or classifying world-wide-web documents into topic hierarchies. We investigate the potential benefits of using a hierarchy over base classes to learn accurate multi-category classifiers for these domains.

First, we consider the possibility of exploiting a class hierarchy as prior knowledge that can help one learn a more accurate classifier. We explore the benefits of learning category-discriminants in a ``hard'' top-down fashion and compare this to a ``soft'' approach which shares training data among sibling categories. Our results show that hierarchies can often help improve prediction accuracy. One reason for this is simply that a hierarchy can help constrain the expressiveness of a hypothesis class in an appropriate manner. However, more interestingly, we also find that using a hierarchy can sometimes help improve prediction accuracy even when controlling for the expressiveness of the hypothesis class.

Return to Greiner's home page