Richard Maclin

NeurIPS Conference 2011 Conference Paper

Advice Refinement in Knowledge-Based SVMs

• Gautam Kunapuli
• Richard Maclin
• Jude Shavlik

Knowledge-based support vector machines (KBSVMs) incorporate advice from domain experts, which can improve generalization significantly. A major limitation that has not been fully addressed occurs when the expert advice is imperfect, which can lead to poorer models. We propose a model that extends KBSVMs and is able to not only learn from data and advice, but also simultaneously improve the advice. The proposed approach is particularly effective for knowledge discovery in domains with few labeled examples. The proposed model contains bilinear constraints, and is solved using two iterative approaches: successive linear programming and a constrained concave-convex approach. Experimental results demonstrate that these algorithms yield useful refinements to expert advice, as well as improve the performance of the learning algorithm overall.

AAAI Conference 2007 Conference Paper

Refining Rules Incorporated into Knowledge-Based Support Vector Learners Via Successive Linear Programming

• Richard Maclin
• Jude Shavlik

Knowledge-based classification and regression methods are especially powerful forms of learning. They allow a system to take advantage of prior domain knowledge supplied either by a human user or another algorithm, combining that knowledge with data to produce accurate models. A limitation of the use of prior knowledge occurs when the provided knowledge is incorrect. Such knowledge likely still contains useful information, but knowledge-based learners might not be able to fully exploit such information. In fact, incorrect knowledge can lead to poorer models than result from knowledge-free learners. We present a support-vector method for incorporating and refining domain knowledge that not only allows the learner to make use of that knowledge, but also suggests changes to the provided knowledge. Our approach is built on the knowledge-based classification and regression methods presented by Fung, Mangasarian, & Shavlik (2002; 2003) and by Mangasarian, Shavlik, & Wild (2004). Experiments on artificial data sets with known properties, as well as on a real-world data set, demonstrate that our method learns more accurate models while also adjusting the provided rules in intuitive ways. Our new algorithm provides an appealing extension to knowledge-based, support-vector learning that is not only able to combine knowledge from rules with data, but is also able to use the data to modify and change those rules to better fit the data.

AAAI Conference 2006 Conference Paper

A Simple and Effective Method for Incorporating Advice into Kernel Methods

• Richard Maclin
• Trevor Walker

We propose a simple mechanism for incorporating advice (prior knowledge), in the form of simple rules, into support-vector methods for both classification and regression. Our approach is based on introducing inequality constraints associated with datapoints that match the advice. These constrained datapoints can be standard examples in the training set, but can also be unlabeled data in a semi-supervised, advice-taking approach. Our new approach is simpler to implement and more efficiently solved than the knowledge-based support vector classification methods of Fung, Mangasarian and Shavlik (2002; 2003) and the knowledge-based support vector regression method of Mangasarian, Shavlik, and Wild (2004), while performing approximately as well as these more complex approaches. Experiments using our new approach on a synthetic task and a reinforcementlearning problem within the RoboCup soccer simulator show that our advice-taking method can significantly outperform a method without advice and perform similarly to prior advice-taking, support-vector machines.

AAAI Conference 2006 Short Paper

Kernel Methods for Word Sense Disambiguation and Acronym Expansion

• Mahesh Joshi
• Richard Maclin

The scarcity of manually labeled data for supervised machine learning methods presents a significant limitation on their ability to acquire knowledge. The use of kernels in Support Vector Machines (SVMs) provides an excellent mechanism to introduce prior knowledge into the SVM learners, such as by using unlabeled text or existing ontologies as additional knowledge sources. Our aim is to develop three kernels – one that makes use of knowledge derived from unlabeled text, the second using semantic knowledge from ontologies, and finally a third, additive kernel consisting of the first two kernels – and study their effect on the tasks of word sense disambiguation and automatic expansion of ambiguous acronyms.

AAAI Conference 2005 Conference Paper

Giving Advice about Preferred Actions to Reinforcement Learners Via Knowledge-Based Kernel Regression

• Richard Maclin
• Lisa Torrey

We present a novel formulation for providing advice to a reinforcement learner that employs supportvector regression as its function approximator. Our new method extends a recent advice-giving technique, called Knowledge-Based Kernel Regression (KBKR), that accepts advice concerning a single action of a reinforcement learner. In KBKR, users can say that in some set of states, an action’s value should be greater than some linear expression of the current state. In our new technique, which we call Preference KBKR (Pref-KBKR), the user can provide advice in a more natural manner by recommending that some action is preferred over another in the specified set of states. Specifying preferences essentially means that users are giving advice about policies rather than Q values, which is a more natural way for humans to present advice. We present the motivation for preference advice and a proof of the correctness of our extension to KBKR. In addition, we show empirical results that our method can make effective use of advice on a novel reinforcement-learning task, based on the RoboCup simulator, which we call Breakaway. Our work demonstrates the significant potential of advice-giving techniques for addressing complex reinforcement learning problems, while further demonstrating the use of support-vector regression for reinforcement learning.

AAAI Conference 1998 Conference Paper

Boosting Classifiers Regionally

• Richard Maclin

This paper presents a newalgorithm for Boosting the performance of an ensembleof classifiers. In Boosting, a series of classifiers is usedto predict the class of data wherelater members of the series concentrate on training data that is incorrectly predicted by earlier members. To makea prediction abouta newpattern, each classifier predicts the class of the pattern and these predictions are then combined. In standard Boosting, the predictions are combined by weightingthe predictions bya termrelated to the accuracyof the classifier on the training data. This approachignores the fact that later classifiers focuson smallsubsets of the patterns and thus mayonly be good at classifying similar patterns. In RegionBoost, this problemis addressed by weighting each classifier’s predictions by a factor measuringhowwell that classifier performson similar patterns. In this paper weexamineseveral methodsfor determininghowwell a classifier performson similar patterns. Empirical tests indicate RegionBoost producesgains in performancefor somedata sets andhas little effect onothers.

AAAI Conference 1997 Conference Paper

An Empirical Evaluation of Bagging and Boosting

• Richard Maclin

An ensemble consists of a set of independently trained classifiers (such as neural networks or decision trees) whose predictions are combined when classifying novel instances. Previous research has shown that an ensemble as a whole is often more accurate than any of the single classifiers in the ensemble. Bagging (Breiman 1996a) and Boosting (F’reund & Schapire 1996) are two relatively new but popular methods for producing ensembles. In this paper we evaluate these methods using both neural networks and decision trees as our classification algorithms. Our results clearly show two important facts. The first is that even though Bagging almost always produces a better classifier than any of its individual component classifiers and is relatively impervious to overfitting, it does not generalize any better than a baseline neural-network ensemble method. The second is that Boosting is a powerful technique that can usually produce better ensembles than Bagging; however, it is more susceptible to noise and can quickly overfit a data set.

IJCAI Conference 1997 Conference Paper

Ensembles as a Sequence of Classifiers

• Lars Asker
• Richard Maclin

An ensemble is a classifier created by combining the predictions of multiple component classifiers. We present a new method for combining classifiers into an ensemble based on a simple estimation of each classifier's competence. The classifiers are grouped into an ordered list where each classifier has a corresponding threshold. To classify an example, the first classifier on the list is consulted and if that classifier's confidence for predicting the example is above the classifier's threshold, then that classifier's prediction is used. Otherwise, the next classifier and its threshold is consulted and so on. If none of the classifiers predicts the example above its confidence threshold then the class of the example is predicted by averaging all of the component classifier predictions. The key to this method is the selection of the confidence threshold for each classifier. We have implemented this method in a system called SEQUEL which has been applied to the task of recognizing volcanos in SAR images of Venus. In this domain, SEQUEL outperforms each individual classifier as well as the simple approach of using an ensemble constructed from the average prediction of all the classifiers.

ICML Conference 1997 Conference Paper

Feature Engineering and Classifier Selection: A Case Study in Venusian Volcano Detection

• Lars Asker
• Richard Maclin

IJCAI Conference 1995 Conference Paper

Combining the Predictions of Multiple Classifiers: Using Competitive Learning to Initialize Neural Networks

• Richard Maclin
• Jude W Shavlik

The primary goal of inductive learning is to generalize well - that is, induce a function that accurately produces the correct output for future inputs. Hansen and Salamon showed that, under certain assumptions, combining the predictions of several separately trained neural networks will improve generalization. One of their key assumptions is that the individual networks should be independent in the errors they produce. In the standard way of performing backpropagation this assumption may be violated, because the standard procedure is to initialize network weights in the region of weight space near the origin. This means that backpropagation's gradient-descent search may only reach a small subset of the possible local minima. In this paper we present an approach to initializing neural networks that uses competitive learning to intelligently create networks that are originally located far from the origin of weight space, thereby potentially increasing the set of reachable local minima. We report experiments on two real-world datasets where combinations of networks initialized with our method generalize better than combinations of networks initialized the traditional way.

AAAI Conference 1994 Conference Paper

Incorporating Advice into Agents that Learn from Reinforcements

• Richard Maclin

Learning from reinforcements is a promising approach for creating intelligent agents. However, reinforcement learning usually requires a large number of training episodes. We present an approach that addresses this shortcoming by allowing a connectionist Q-learner to accept advice given, at any time and in a natural manner, by an external observer. In our approach, the advice-giver watches the learner and occasionally makes suggestions, expressed as instructions in a simple programming language. Based on techniques from knowledge-based neural networks, these programs are inserted directly into the agent’ s utility function. Subsequent reinforcement learning further integrates and refines the advice. We present empirical evidence that shows our approach leads to statistically-significant gains in expected reward. Importantly, the advice improves the expected reward regardless of the stage of training at which it is given.

AAAI Conference 1992 Conference Paper

Using Knowledge-Based Neural Networks to Improve Algorithms: Refining the Chou-Fasman Algorithm for Protein Folding

• Richard Maclin

We describe a method for using machine learning to refine algorithms represented as generalized finite-state automata. The knowledge in an automaton is translated into an artificial neural network, and then refined with backpropagation on a set of examples. Our technique for translating an automaton into a network extends KBANN, a system that translates a set of propositional rules into a corresponding neural network. The extended system, FSKBANN, allows one to refine the large class of algorithms that can be represented as state-based processes. As a test, we use FSKBANN to refine the Chou-Fasman algorithm, a method for predicting how globular proteins fold. Empirical evidence shows the refined algorithm FSKBANN produces is statistically significantly more accurate than both the original Chou-Fasman algorithm and a neural network trained using the standard approach.