Faster Hoeffding Racing: Bernstein Races via Jackknife Estimates

Hoeffding racing algorithms are used to achieve computational speedups in settings where the goal is to select a “best” option among a set of alternatives, but the amount of data is so massive that scoring all candidates using every data point is too costly. The key is to construct confidence intervals for scores of candidates that are used to eliminate options sequentially as more samples are processed. We propose a tighter version of Hoeffding racing based on empirical Bernstein inequalities, where a jackknife estimate is used in place of the unknown variance. We provide rigorous proofs of the accuracy of our confidence intervals in the case of U-statistics and entropy estimators, and demonstrate the efficacy of our racing algorithms with synthetic experiments.

[1]  Devdatt P. Dubhashi,et al.  Concentration of Measure for the Analysis of Randomized Algorithms: Contents , 2009 .

[2]  R. Serfling.  Approximation Theorems of Mathematical Statistics , 1980 .

[3]  P. Rousseeuw,et al.  Wiley Series in Probability and Mathematical Statistics , 2005 .

[4]  Csaba Szepesvári,et al.  Empirical Bernstein stopping , 2008, ICML '08.

[5]  Geoff Hulten,et al.  Mining high-speed data streams , 2000, KDD '00.

[6]  A. Antos,et al.  Convergence properties of functional estimates for discrete distributions , 2001 .

[7]  Andrew W. Moore,et al.  Hoeffding Races: Accelerating Model Selection Search for Classification and Function Approximation , 1993, NIPS.

[8]  H. Robbins.  Some aspects of the sequential design of experiments , 1952 .

[9]  Philip S. Yu,et al.  Distributed hoeffding trees for pocket data mining , 2011, 2011 International Conference on High Performance Computing & Simulation.

[10]  Alan J. Lee,et al.  U-Statistics: Theory and Practice , 1990 .

[11]  Bernard Zenko,et al.  Speeding-Up Hoeffding-Based Regression Trees With Options , 2011, ICML.

[12]  M. A. Arcones,et al.  A Bernstein-type inequality for U-statistics and U-processes , 1995 .

[13]  A E Bostwick,et al.  THE THEORY OF PROBABILITIES. , 1896, Science.

[14]  W. Hoeffding.  Probability Inequalities for sums of Bounded Random Variables , 1963 .

[15]  Colin McDiarmid,et al.  Surveys in Combinatorics, 1989: On the method of bounded differences , 1989 .

[16]  Liva Ralaivola,et al.  Empirical Bernstein Inequalities for U-Statistics , 2010, NIPS.

[17]  Csaba Szepesvári,et al.  Exploration-exploitation tradeoff using variance estimates in multi-armed bandits , 2009, Theor. Comput. Sci..

[18]  J. Steele.  An Efron-Stein inequality for nonsymmetric statistics , 1986 .

[19]  Geoff Holmes,et al.  New Options for Hoeffding Trees , 2007, Australian Conference on Artificial Intelligence.

[20]  B. Efron,et al.  The Jackknife Estimate of Variance , 1981 .

[21]  Csaba Szepesvári,et al.  Tuning Bandit Algorithms in Stochastic Environments , 2007, ALT.

[22]  A. Dasgupta.  Asymptotic Theory of Statistics and Probability , 2008 .

[23]  R. Serfling.  Approximation Theorems of Mathematical Statistics , 1980 .

[24]  Anne Lohrli.  Chapman and Hall , 1985 .

[25]  P. MassartLedoux,et al.  Concentration Inequalities Using the Entropy Method , 2002 .

[26]  Liam Paninski,et al.  Estimation of Entropy and Mutual Information , 2003, Neural Computation.

[27]  Ruoming Jin,et al.  Efficient decision tree construction on streaming data , 2003, KDD '03.

[28]  Shie Mannor,et al.  Action Elimination and Stopping Conditions for the Multi-Armed Bandit and Reinforcement Learning Problems , 2006, J. Mach. Learn. Res..