Search

We present a simple algorithm for playing a repeated game. We show that a player using this algorithm suffers average loss that is guaranteed to come close to the minimum loss achievable by any fixed strategy. Our bounds are nonasymptotic and hold for any opponent. The algorithm, which uses the multiplicative-weight methods of Littlestone and Warmuth, is analyzed using the Kullback–Liebler divergence. This analysis yields a new, simple proof of the min–max theorem, as well as a provable method of approximately solving a game. A variant of our game-playing algorithm is proved to be optimal in a very strong sense. Journal of Economic Literature Classification Numbers: C44, C70, D83.

We analyze algorithms that predict a binary value by combining the predictions of several prediction strategies, called experts . Our analysis is for worst-case situations, i.e., we make no assumptions about the way the sequence of bits to be predicted is generated. We measure the performance of the algorithm by the difference between the expected number of mistakes it makes on the bit sequence and the expected number of mistakes made by the best expert on this sequence, where the expectation is taken with respect to the randomization in the predictins. We show that the minimum achievable difference is on the order of the square root of the number of mistakes of the best expert, and we give efficient algorithms that achieve this. Our upper and lower bounds have matching leading constants in most cases. We then show how this leads to certain kinds of pattern recognition/learning algorithms with performance bounds that improve on the best results currently know in this context. We also compare our analysis to the case in which log loss is used instead of the expected number of mistakes.

We study the problem of deterministically predicting boolean values by combining the boolean predictions of several experts. Previous on-line algorithms for this problem predict with the weighted majority of the experts'' predictions. These algorithms give each expert an exponential weight beta^m where beta is a constant in [0,1) and m is the number of mistakes made by the expert in the past. We show that it is better to use sums of binomials as weights. In particular, we present a deterministic algorithm using binomial weights that has a better worst case mistake bound than the best deterministic algorithm using exponential weights. The binomial weights naturally arise from a version space argument. We also show how both exponential and binomial weighting schemes can be used to make prediction algorithms robust against noise.

The Fluidic Lens Camera System provides an exciting opportunity for the Image Processing Community. Designed for a surgical environment, this camera has higher magnification and has better portability than traditional laparoscopic cameras. From an image processing prospective, the fluid causes non-uniform blur of different color planes. While the green image is sharp, the red and blue images are blurred. Previous methods have been developed to separate out the edge and shading components of the green image and to use the edge information in green to replace the blurred blue edges. This algorithm succeed in most areas, however in some areas, color bleeding artifacts occurred. We restate this problem as a classification problem. Using the contourlet and wavelet coefficients as features, the proposed algorithm determines in what areas color bleeding will occur and does not apply the sharpening algorithm in these areas. By applying the previous contourlet method in areas where it succeeds, we can produce an overall sharper image with reduced color bleeding artifacts. The ability to correctly classify when the previous algorithm will succeed is crucial to the success of the algorithm. The principal application is medical imaging, however, the fields of satellite pan-sharpening and image denoising can benefit from the results found in this paper.

We present a simple variant of the k-d tree which automatically adapts to intrinsic low dimensional structure in data without having to explicitly learn this structure.

A simple and computationally efficient scheme for tree-structured vector quantization is presented. Unlike previous methods, its quantization error depends only on the intrinsic dimension of the data distribution, rather than the apparent dimension of the space in which the data happen to lie.

We introduce novel profile-based string kernels for use with support vector machines (SVMs) for the problems of protein classification and remote homology detection. These kernels use probabilistic profiles, such as those produced by the PSI-BLAST algorithm, to define position-dependent mutation neighborhoods along protein sequences for inexact matching of k-length subsequences ("k-mers") in the data. By use of an efficient data structure, the kernels are fast to compute once the profiles have been obtained. For example, the time needed to run PSI-BLAST in order to build the profiles is significantly longer than both the kernel computation time and the SVM training time. We present remote homology detection experiments based on the SCOP database where we show that profile-based string kernels used with SVM classifiers strongly outperform all recently presented supervised SVM methods. We further examine how to incorporate predicted secondary structure information into the profile kernel to obtain a small but significant performance improvement. We also show how we can use the learned SVM classifier to extract "discriminative sequence motifs" — short regions of the original profile that contribute almost all the weight of the SVM classification score — and show that these discriminative motifs correspond to meaningful structural features in the protein data. The use of PSI-BLAST profiles can be seen as a semi-supervised learning technique, since PSI-BLAST leverages unlabeled data from a large sequence database to build more informative profiles. Recently presented "cluster kernels" give general semi-supervised methods for improving SVM protein classification performance. We show that our profile kernel results also outperform cluster kernels while providing much better scalability to large datasets. Supplementary website:.

We study a simple learning algorithm for binary classification. Instead of predicting with the best hypothesis in the hypothesis class, that is, the hypothesis that minimizes the training error, our algorithm predicts with a weighted average of all hypotheses, weighted exponentially with respect to their training error. We show that the prediction of this algorithm is much more stable than the prediction of an algorithm that predicts with the best hypothesis. By allowing the algorithm to abstain from predicting on some examples, we show that the predictions it makes when it does not abstain are very reliable. Finally, we show that the probability that the algorithm abstains is comparable to the generalization error of the best hypothesis in the class., Published by the Institute of Mathematical Statistics (http://www.imstat.org) in the Annals of Statistics (http://www.imstat.org/aos/) at http://dx.doi.org/10.1214/009053604000000058

Discussions of: "Process consistency for AdaBoost" [Ann. Statist. 32 (2004), no. 1, 13-29] by W. Jiang; "On the Bayes-risk consistency of regularized boosting methods" [ibid., 30-55] by G. Lugosi and N. Vayatis; and "Statistical behavior and consistency of classification methods based on convex risk minimization" [ibid., 56-85] by T. Zhang. Includes rejoinders by the authors.

In the multi-armed bandit problem, a gambler must decide which arm of K non-identical slot machines to play in a sequence of trials so as to maximize his reward. This classical problem has received much attention because of the simple model it provides of the trade-off between exploration (trying out each arm to find the best one) and exploitation (playing the arm believed to give the best payoff). Past solutions for the bandit problem have almost always relied on assumptions about the statistics of the slot machines. In this work, we make no statistical assumptions whatsoever about the nature of the process generating the payoffs of the slot machines. We give a solution to the bandit problem in which an adversary, rather than a well-behaved stochastic process, has complete control over the payoffs. In a sequence of T plays, we prove that the expected per-round payoff of our algorithm approaches that of the best arm at the rate O(T/sup -1/3/), and we give an improved rate of convergence when the best arm has fairly low payoff. We also consider a setting in which the player has a team of "experts" advising him on which arm to play; here, we give a strategy that will guarantee expected payoff close to that of the best expert. Finally, we apply our result to the problem of learning to play an unknown repeated matrix game against an all-powerful adversary.

We examine the problem of learning to play various games optimally against resource-bounded adversaries, with an explicit emphasis on the computational efficiency of the learning algorithm. We are especially interested in providing efficient algorithms for games other than penny-matching (in which payoff is received for matching the adversary's action in the current round), and for adversaries other than the classically studied finite automata. In particular, we examine games and adversaries for which the learning algorithm's past actions may strongly affect the adversary's future willingness to "cooperate" (that is, permit high payoff), and therefore require carefully planned actions on the part of the learning algorithm. For example, in the game we call contract, both sides play O or 1 on each round, but our side receives payoff only if we play 1 in synchrony with the adversary; unlike penny-matching, playing O in synchrony with the adversary pays nothing. The name of the game is derived from the example of signing a contract, which becomes valid only if both parties sign (play 1).

One of the surprising recurring phenomena observed in experiments with boosting is that the test error of the generated classifier usually does not increase as its size becomes very large, and often is observed to decrease even after the training error reaches zero. In this paper, we show that this phenomenon is related to the distribution of margins of the training examples with respect to the generated voting classification rule, where the margin of an example is simply the difference between the number of correct votes and the maximum number of votes received by any incorrect label. We show that techniques used in the analysis of Vapnik’s support vector classifiers and of neural networks with small weights can be applied to voting methods to relate the margin distribution to the test error. We also show theoretically and experimentally that boosting is especially effective at increasing the margins of the training examples. Finally, we compare our explanation to those based on the bias-variance decomposition.

We introduce and analyze a new algorithm for linear classification which combines Rosenblatt‘s perceptron algorithm with Helmbold and Warmuth‘s leave-one-out method. Like Vapnik‘s maximal-margin classifier, our algorithm takes advantage of data that are linearly separable with large margins. Compared to Vapnik‘s algorithm, however, ours is much simpler to implement, and much more efficient in terms of computation time. We also show that our algorithm can be efficiently used in very high dimensional spaces using kernel functions. We performed some experiments using our algorithm, and some variants of it, for classifying images of handwritten digits. The performance of our algorithm is close to, but not as good as, the performance of maximal-margin classifiers on the same problem, while saving significantly on computation time and programming effort.

We study the close connections between game theory, on-line prediction and boosting. After a brief review of game theory, we describe an algorithm for learning to play repeated games based on the on-line prediction methods of Littlestone and Warmuth. The analysis of this algorithm yields a simple proof of von Neumann’s famous minmax theorem, as well as a provable method of approximately solving a game. We then show that the on-line prediction model is obtained by applying this gameplaying algorithm to an appropriate choice of game and that boosting is obtained by applying the same algorithm to the “dual” of this game.

This paper describes new and efficient algorithms for learning deterministic finite automata. Our approach is primarily distinguished by two features: (1) the adoption of an average-case setting to model the “typical” labeling of a finite automaton, while retaining a worst-case model for the underlying graph of the automaton, along with (2) a learning model in which the learner is not provided with the means to experiment with the machine, but rather must learn solely by observing the automaton's output behavior on a random input sequence. The main contribution of this paper is in presenting the first efficient algorithms for learning non-trivial classes of automata in an entirely passive learning model. We adopt an on-line learning model in which the learner is asked to predict the output of the next state, given the next symbol of the random input sequence; the goal of the learner is to make as few prediction mistakes as possible. Assuming the learner has a means of resetting the target machine to a fixed start state, we first present an efficient algorithm that makes an expected polynomial number of mistakes in this model. Next, we show how this first algorithm can be used as a subroutine by a second algorithm that also makes a polynomial number of mistakes even in the absence of a reset. Along the way, we prove a number of combinatorial results for randomly labeled automata. We also show that the labeling of the states and the bits of the input sequence need not be truly random, but merely semi - random . Finally, we discuss an extension of our results to a model in which automata are used to represent distributions over binary strings.

In this work we present some improvements and extensions to previous work on boosting weak learners [Sch90, Fre90]. Our main result is an improvement of the boosting-by-majority algorithm. One implication of the performance of this algorithm is that if a concept class can be learned in the PAC model to within some fixed error smaller than 1/2, then it can be learned to within an arbitrarily small error e > 0 with time complexity 0((1/e)(log 1/e)2) (fixing the sample space and concept class and the required reliability). We show that the majority rule is the optimal rule for combining general weak learners. We also extend the boosting algorithm to concept classes that give multi-valued labels and real-valued labels.

We combine the results of [13] and [8] and derive a continuous variant of a large class of drifting games. Our analysis furthers the understanding of the relationship between boosting, drifting games, and Brownian motion and yields a differential equation that describes the core of the problem.

We apply the exponential weight algorithm, introduced and Littlestone and Warmuth [26] and by Vovk [35] to the problem of predicting a binary sequence almost as well as the best biased coin. We first show that for the case of the logarithmic loss, the derived algorithm is equivalent to the Bayes algorithm with Jeffrey’s prior, that was studied by Xie and Barron [38] under probabilistic assumptions. We derive a uniform bound on the regret which holds for any sequence. We also show that if the empirical distribution of the sequence is bounded away from 0 and from 1, then, as the length of the sequence increases to infinity, the difference between this bound and a corresponding bound on the average case regret of the same algorithm (which is asymptotically optimal in that case) is only 1/2. We show that this gap of 1/2 is necessary by calculating the regret of the min–max optimal algorithm for this problem and showing that the asymptotic upper bound is tight. We also study the application of this algorithm to the square loss and show that the algorithm that is derived in this case is different from the Bayes algorithm and is better than it for prediction in the worst-case.

In the multiarmed bandit problem, a gambler must decide which arm of K nonidentical slot machines to play in a sequence of trials so as to maximize his reward. This classical problem has received much attention because of the simple model it provides of the trade-off between exploration (trying out each arm to find the best one) and exploitation (playing the arm believed to give the best payoff). Past solutions for the bandit problem have almost always relied on assumptions about the statistics of the slot machines. In this work, we make no statistical assumptions whatsoever about the nature of the process generating the payoffs of the slot machines. We give a solution to the bandit problem in which an adversary, rather than a well-behaved stochastic process, has complete control over the payoffs. In a sequence of T plays, we prove that the per-round payoff of our algorithm approaches that of the best arm at the rate O(T-1/2). We show by a matching lower bound that this is the best possible. We also prove that our algorithm approaches the per-round payoff of any set of strategies at a similar rate: if the best strategy is chosen from a pool of N strategies, then our algorithm approaches the per-round payoff of the strategy at the rate O((log N1/2 T-1/2). Finally, we apply our results to the problem of playing an unknown repeated matrix game. We show that our algorithm approaches the minimax payoff of the unknown game at the rate O(T-1/2).