Arrow Research search

Author name cluster

Foster John Provost

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

3 papers
1 author row

Possible papers

3

AAAI Conference 1996 Conference Paper

Scaling Up: Distributed Machine Learning with Cooperation

  • Foster John Provost

Machine-learning methods are becoming increasingly popular for automated data analysis. However, standard methods do not scale up to massive scientific and business data sets without expensive hardware. This paper investigates a practical alternative for scaling up: the use of distributed processing to take advantage of the often dormant PCs and workstations available on local networks. Each workstation runs a common rule-learning program on a subset of the data. We first show that for commonly used ruleevaluation criteria, a simple form of cooperation can guarantee that a rule will look good to the set of cooperating learners if and only if it would look good to a single learner operating with the entire data set. We then show how such a system can further capitalize on different perspectives by sharing learned knowledge for significant reduction in search effort. We demonstrate the power of the method by learning from a massive data set taken from the domain of cellular fraud detection. Finally, we provide an overview of other methods for scaling up machine learning. heterogeneous workstations. We use a standard rulelearning algorithm, modified slightly to allow cooperation between learners. At a high level, our metaphor for distributed learning is one of cooperating experts, each of which has a slightly different perspective on the concept to be learned. We define cooperation as the learning-time sharing of information to increase the quality of the learned knowledge or to reduce or redirect the search. The learners communicate with each other by passing messages. The group can take advantage of the communication by asking questions or by sharing learned knowledge.

AAAI Conference 1993 Conference Paper

Iterative Weakening: Optimal and Near-Optimal Policies for the Selection of Search Bias

  • Foster John Provost

Decisions made in setting up and running search programs bias the searches that they perform. Search bias refers to the definition of a search space and the definition of the program that navigates the space. This paper addresses the problem of using knowledge regarding the complexity of various syntactic search biases to form a policy for selecting bias. In particular, this paper shows that a simple policy, iterative weakening, is optimal or nearly optimal in cases where the biases can be ordered by computational complexity and certain relationships hold between the complexity of the various biases. The results are obtained by viewing bias selection as a (higher-level) search problem. Iterative weakening evaluates the states in order of increasing complexity. An offshoot of this work is the formation of a near-optimal policy for selecting both breadth and depth bounds for depth-fist search with very large (possibly unbounded) breadth and depth.

AAAI Conference 1992 Conference Paper

Inductive Policy

  • Foster John Provost

The concept of inductive bias can be broken down into the underlying assumptions of the domain, the particular implementation choices that restrict or order the space of hypotheses considered by the learning program (the bias choices), and the inductive policy that links the two. We define inductive policy as the strategy used to make bias choices based on the underlying assumptions. Inductive policy decisions involve addressing tradeoffs with respect to different bias choices. Without addressing these tradeoffs, bias choices will be made arbitrarily. From the standpoint of inductive policy, we discuss two issues not addressed much in the machine learning literature. First we discuss batch learning with a strict time constraint, and present an initial study with respect to trading off predictive accuracy for speed of learning. Next we discuss the issue of learning in a domain where different types of errors have different associated costs (risks). We show that by using different inductive policies accuracy can be traded off for safety. We also show how the value for the latter tradeoff can be represented explicitly in a system that adjusts bias choices with respect to a particular inductive policy.

v2026.09.13