Showing papers on "Algorithmic learning theory published in 1981"

Machine learning research

Abstract: META is a Natural Language learning project whose specific objective is to acquire new word definitions and new concepts from contextual information in interactive dialogues. It is an instance of a learning-from-examples method, with a difference: learning proceeds in a reactive, knowledge-rich environment. Our initial research indicates that the interactive nature of the environment ought to be a crucial component of any general learning system. In brief, our investigations have led us to formulate the following hypotheses, which we intend to test and pursue further in the immediate future:1. Learning requires progressive refinement -- It is unreasonable to expect that a computer system (or a human) learn a concept or a skill without error, in its full embellished form, in one brief learning session. It must undergo a sequence of progressive test-and-update stages. In other words, a concept can be learned by first inducing a rough approximation of its final form and successively correcting this approximation with more accurate, more detailed information.2. Interaction with a reactive environment -- Interaction is the engine that drives learning processes. A learner must be able to direct queries to its teacher or perform experiments on its environment. It must be able test out new concepts and skills as it learns them.3. Reasoning by Analogy -- The more a system can learn by relating new concepts to old, by modifying existing concepts, or using chunks of existing concepts as building blocks, the more robust and general its learning mechanisms will be.

Learning Concepts by Performing Experiments

Abstract: Marvin is a program which is capable of learning concepts from many different environments. It achieves this by using a flexible description language based on first order predicate logic with quantifiers. Once a concept has been learnt, Marvin treats the concept description as a program which can be executed to produce an output. Thus the learning system can also be viewed as an automatic program synthesizer. The ability to treat a concept as a program permits the learning system to construct objects to show a human trainer. Given an initial example by the trainer, Marvin creates a concept intended to describe the class of objects containing the example. The validity of the description is tested when Marvin constructs an instance of the concept to show the trainer. If he indicates that the example constructed by the program belongs to the concept which is to be learnt, called the 'target', then Marvin attempts to generalize the description of its hypothesized concept. If the example does not belong to the target then the description must be made more specific so that a correct example can be constructed. This process is repeated until the description of the concept cannot be generalized without producing unacceptable examples. Marvin has an associative memory which enables it to match the descriptions of objects it is shown with concepts that it has stored in memory. Complex concepts are learnt by first learning the descriptions of simple concepts which provide Marvin with the knowledge necessary to learn the more complex ones. A concept may represent a non-deterministic program, that is, more than one output may result from the same input. Not all the possible outputs of a concept are acceptable as training instances. Thus, Marvin must have an 'instance selector' which is capable is choosing the best objects to show the trainer. Marvin has been tested on a number of learning tasks. Extensive performance measurements were made during these sessions with the program. The results indicate that Marvin is capable of learning complex concepts quite quickly.

Machine Learning for Information Management

Abstract: : This paper discusses machine learning in the context of information management. The core idea is that of a compiler system that can hold a conversation with a user in English about his specific domain of interest, subsequently retrieve and display information conveyed by the user, and apply various types of external software systems to solve user problems. The specific learning problem discussed is how to enable computer systems to acquire information about domains with which they are unfamiliar from people who are expert in those domains, but have little or no training in computer science. The information to be acquired is that needed to support question-answering or fact retrieval tasks, and the type of learning to be employed is learning by being told. Reflecting the intimate connection between language and reasoning, this paper is largely concerned with the problems of learning concepts and language simultaneously.

On the Algorithmic Theory of Dictionaries

Abstract: A class of algebraic structures called dictionaries is defined. Among the properties (axioms) of dictionaries we find the algorithmic property : for every s the program while empty(s) do s := delete(amember(s),s) terminates. Starting with this observation we define and develop the algorithmic theory of dictionaries ATD. We are proving the representation theorem: every dictionary structure is isomorphic to the family of all finite subsets of some set. The complexity of the set of theorems of various extensions of ATD can vary from a hiperarithmetical set to the complement of a recursively enumerable set.