Query answering usually assumes that the asker is looking for a single correct answer to the question. When retrieving a textual answer this is often the case, but when searching for numeric answers, there are additional considerations. In particular, numbers often have units associated with them, and the asker may not care whether the raw answer is in feet or meters. Also, numbers usually denote a precision. In a few cases, the precision may be explicit, but normally, there is an implied precision associated with every number. Finally, an association between different reliability levels to different sources can be made. In this paper, the authors experimentally show that, in the context of conflicting answers from multiple sources, numeric query accuracy can be improved by taking advantage of units, precision, and the reliability of sources.

AdaBoost.M2 and AdaBoost.MH are boosting algorithms for learning from multiclass datasets. They have received less attention than other boosting algorithms because they require base classifiers that can handle the pseudoloss or Hamming loss, respectively. The difficulty with these loss functions is that each example is associated with k weights, where k is the number of classes. We address this issue by transforming an m-example dataset with k weights per example into a dataset with km examples and one weight per example. Minimising error on the transformed dataset is equivalent to minimising loss on the original dataset. Resampling the transformed dataset can be used for time efficiency and base classifiers that cannot handle weighted examples. We empirically apply the transformation on several multiclass datasets using naive Bayes and decision trees as base classifiers. Our experiment shows that it is competitive with AdaBoost.ECC, a boosting algorithm using output coding.

Abstract: This paper describes an extension of the traditional application of Genetic Programming in the domain of the prediction of daily currency exchange rates. In combination with trigonometric operators, we introduce a new set of high-order statistical functions in a unique representation and analyze each system performance using daily returns of the British Pound and Japanese Yen. We will demonstrate that the introduction of high-order statistical functions in combination with trigonometric functions will outperform other traditional models such as Genetic Programming with the basic function set and ARMA models. Performance will be measured on hit percentage, average percentage change, and profit.

A linear threshold function (LTF) is probabilistically consistent (p-consistent) on a noisy distribution of labeled examples if the probability of noise decreases monotonically with the absolute difference from the LTF's threshold. I show that pconsistent LTFs are PAC-learnable in polynomial time if the separation between examples and the LTF is sufficiently large, and if the vector length of each example is sufficiently small. This result generalizes previous polynomial learnability results for binary independent features (often called naive Bayes), linear least squares, and classification noise. However, the bound derived for the number of examples is not a low-order polynomial, and multiple iterations are required with one update per iteration (any incremental algorithm that PAC-learns LTFs can be utilized for updating) . Nevertheless, the algorithm performs well on commonly-used datasets. 1 INTRODUCTION The problem of learning from examples is usually complicated by noise. A given...

This paper develops and analyzes a new online algorithm for learning linear functions, called the Binary Exponentiated Gradient algorithm (BEG).BEG imposes an lower and upper bound for all the weights.Using Kivinen and Warmuth's methodology, the BEG algorithm is developed from a binary entropy distance function and the square loss function, and worst-case upper bounds on the square loss are demonstrated for BEG on arbitrary sequences of trials (instance-outcome pairs).BEG's behavior is unusual in that in some situations its worst-case behavior is comparable to the wellknown gradient descent algorithms, e.g.Widrow-Hoff, while in others, it is comparable to the newer exponentiated gradient algorithms.An experiment shows when it outperforms both algorithms.

We demonstrate sufficient conditions for polynomial learnability of suboptimal linear threshold functions using perceptrons. The central result is as follows. Suppose there exists a vector w*, of n weights (including the threshold) with “accuracy” 1 − α, “average error” η, and “balancing separation” σ, i.e., with probability 1 − α, w* correctly classifies an example x; over examples incorrectly classified by w*, the expected value of |w* · x| is η (source of inaccuracy does not matter); and over a certain portion of correctly classified examples, the expected value of |w* · x| is σ. Then, with probability 1 − δ, the perceptron achieves accuracy at least 1 − [∊ + α(1 + η/σ)] after O[n∊−2σ−2(ln 1/δ)] examples.

I show that the linear threshold functions are polynomially learnable in the presence of classification noise, i.e., polynomial in n, 1/ε, 1/δ, and 1/σ, where n is the number of Boolean attributes, ε and δ are the usual accuracy and confidence parameters, and σ indicates the minimum distance of any example from the target hyperplane, which is assumed to be lower than the average distance of the examples from any hyperplane. This result is achieved by modifying the Perceptron algorithm—for each update, a weighted average of a sample of misclassified examples and a correction vector is added to the current weight vector. Similar modifications are shown for the Weighted Majority algorithm. The correction vector is simply the mean of the normalized examples. In the special case of Boolean threshold functions, the modified Perceptron algorithm performs O (n2ε−2 ) iterations over O(n4ε −2ln(n/(δε))) examples. This improves on the previous classification-noise result of Angluin and Laird to a much larger concept class with a similar number of examples, but with multiple iterations over the examples.

I demonstrate sufficient conditions for polynomial learnability of suboptimal linear threshold functions. The central result is as follows. Suppose there exists a vector ~ w of n weights (including the threshold) with "accuracy " 1 \\Gamma ff, "average error" j, and "balancing separation" oe, i.e., with probability 1 \\Gamma ff, ~ w correctly classifies an example ~x; over examples incorrectly classified by ~ w , the expected value of j ~ w \\Delta ~xj is j (source of inaccuracy does not matter); and over a certain portion of correctly classified examples, the expected value of j ~ w \\Delta ~xj is oe. Then, both the perceptron and weighted majority algorithms achieve an online accuracy (i.e., accuracy over all examples) at least 1 \\Gamma (ffl + ff(1 + j=oe)) with confidence at least 1 \\Gamma ffi in time and number of examples polynomial in 1=oe, 1=ffl, 1=ffi, and n. 1 Introduction Recently, it has been shown that learning probably approximately optimal (PAO) linear-threshold funct...

This paper argues that tractability should be a primary concern for computational theories of intelligence. The majority of research in artificial intelligence is devoted to the study of intractable problems and formalisms, but these do not directly address tractability. This paper suggest that one methodology for discovering tractable problems is to construct and analyse ad hoc programs.

Structured matching is a task-specific technique for selecting one choice out of a small number of alternatives based on a given set of parameters. In structured matching, the knowledge and control for making a decision are integrated within a hierarchical structure. Each node in the hierarchy corresponds to a different aspect of the decision and contains knowledge for directly mapping the results of its children nodes (or selected parameters) into a choice on the subdecision. The root node selects the final choice for the decision. The authors formally characterize the task and strategy of structured matching and analyze its computational complexity. They believe that structured matching captures the essence of what makes a range of decision-making problems computationally feasible to solve.>

To understand commonsense reasoning, it is necessary to discover what kinds of problems a commonsense reasoner should be able to solve, what the reasoner needs to have in order to solve those problems, and the relationships among the various kinds of problem-solving abilities. Four methods for performing qualitative reasoning about the behaviour of physical situations are examined. Three of the methods perform qualitative simulation, which determines the behavior of a situation by a qualitative version of simulation methods. The other method is called consolidation, which derives the behavior of a situation by composing the behavior of the situation's components. The work shows that qualitative simulation and consolidation work on different problems of qualitative reasoning, and that their differences and similarities lead to several implications about their role in qualitative reasoning.>