Results 1  10
of
13
A Racing Algorithm for Configuring Metaheuristics
, 2002
"... This paper describes a racing procedure for finding, in a limited amount of time, a configuration of a metaheuristic that performs as good as possible on a given instance class of a combinatorial optimization problem. Taking inspiration from methods proposed in the machine learning literature ..."
Abstract

Cited by 117 (34 self)
 Add to MetaCart
This paper describes a racing procedure for finding, in a limited amount of time, a configuration of a metaheuristic that performs as good as possible on a given instance class of a combinatorial optimization problem. Taking inspiration from methods proposed in the machine learning literature for model selection through crossvalidation, we propose a procedure that empirically evaluates a set of candidate configurations by discarding bad ones as soon as statistically sufficient evidence is gathered against them. We empirically evaluate our procedure using as an example the configuration of an ant colony optimization algorithm applied to the traveling salesman problem.
Adaptive problemsolving for largescale scheduling problems: A case study
, 1996
"... Although most scheduling problems are NPhard, domain specific techniques perform well in practice but are quite expensive to construct. In adaptive problemsolving, domain specific knowledge is acquired automatically for a general problem solver with a flexible control architecture. In this approac ..."
Abstract

Cited by 25 (3 self)
 Add to MetaCart
Although most scheduling problems are NPhard, domain specific techniques perform well in practice but are quite expensive to construct. In adaptive problemsolving, domain specific knowledge is acquired automatically for a general problem solver with a flexible control architecture. In this approach, a learning system explores a space of possible heuristic methods for one wellsuited to the eccentricities of the given domain and problem distribution. In this article, we discuss an application of the approach to scheduling satellite communications. Using problem distributions based on actual mission requirements, our approach identifies strategies that not only decrease the amount of CPU time required to produce schedules, but also increase the percentage of problems that are solvable within computational resource limitations. 1.
Distortioninvariant recognition via jittered queries
 In Computer Vision and Pattern Recognition (CVPR2000
, 2000
"... This paper presents a new approach for achieving distortioninvariant recognition and classification. A test example to be classified is viewed as a query intended to find similar examples in the training set (or class models derived from the training set). The key idea is that instead of querying w ..."
Abstract

Cited by 16 (3 self)
 Add to MetaCart
This paper presents a new approach for achieving distortioninvariant recognition and classification. A test example to be classified is viewed as a query intended to find similar examples in the training set (or class models derived from the training set). The key idea is that instead of querying with a single pattern, we construct a more robust query, based on the family of patterns formed by distorting the test example. Although query execution is slower than if the invariances were successfully precompiled during training, there are significant advantages in several contexts: (i) providing invariances in memorybased learning, (ii) in model selection, where reducing training time at the expense of test time is a desirable tradeoff, and (iii) in enabling robust, ad hoc searches based on a single example. Preliminary tests for memorybased learning on the NIST handwritten digit database with a limited set of shearing and translation distortions produced an error rate of 1.35%. 1
Sequential Inductive Learning
 In Proceedings of the Thirteenth National Conference on Artificial Intelligence
, 1995
"... In this paper I advocate a new model for inductive learning. Called sequential induction, this model bridges classical fixedsample learning techniques (which are efficient but ad hoc), and worstcase approaches (which provide strong statistical guarantees but are too inefficient for practical use). ..."
Abstract

Cited by 7 (0 self)
 Add to MetaCart
In this paper I advocate a new model for inductive learning. Called sequential induction, this model bridges classical fixedsample learning techniques (which are efficient but ad hoc), and worstcase approaches (which provide strong statistical guarantees but are too inefficient for practical use). According to the sequential inductive model, learning is a sequence of decisions which are informed by training data. By analyzing induction at the level of these decisions, and by utilizing the minimum data necessary to make each decision, sequential inductive techniques can provide the strong statistical guarantees of worstcase methods, but with substantially less data than those methods require. The sequential inductive model is also useful as a method for determining a sufficient sample size for inductive learning and as such, is relevant to megainduction,where the preponderance of data introduces problems of scale. The peepholing and decisiontheoretic subsampling approaches of Catlet...
Automating the Process of Optimization in Spacecraft Design
 IEEE AEROSPACE CONFERENCE PROCEEDINGS
, 1997
"... Spacecraft design optimization is a difficult problem, due to the complexity of optimization cost surfaces and the human expertise in optimization that is necessary in order to achieve good results. In this paper, we propose the use of a set of generic, metaheuristic optimization algorithms (e.g., g ..."
Abstract

Cited by 6 (2 self)
 Add to MetaCart
Spacecraft design optimization is a difficult problem, due to the complexity of optimization cost surfaces and the human expertise in optimization that is necessary in order to achieve good results. In this paper, we propose the use of a set of generic, metaheuristic optimization algorithms (e.g., genetic algorithms, simulated annealing), which is configured for a particular optimization problem by an adaptive problem solver based on artificial intelligence and machine learning techniques. We describe work in progress on OASIS, a system for adaptive problem solving based on these principles.
An Effective Method for Correlated Selection Problems
, 1994
"... This article is organized as follows. The next selection reviews the standard statistical approaches to selection problems and introduce some of the terminology that has been developed in this area. Section 3 discusses the method of multiple comparisons and illustrate how this method can be applied ..."
Abstract

Cited by 4 (2 self)
 Add to MetaCart
This article is organized as follows. The next selection reviews the standard statistical approaches to selection problems and introduce some of the terminology that has been developed in this area. Section 3 discusses the method of multiple comparisons and illustrate how this method can be applied to solving correlated selection problems. From this perspective, a selection problem reduces to the problem of simultaneously performing a number of twohypothesis selection problems. Section 4 describes the Sequential Probability Ratio Test (SPRT), an efficient method for solving a twohypothesis selection problem. Section 5 describes a new correlated selection method, MCSPRT, which combines the multiple comparison approach with the SPRT. Section 6 shows how the cost of selection can be further improved by a decisiontheoretic evaluation of the cost of taking examples.
Efficient heuristic hypothesis ranking
 Journal of Artificial Intelligence Research
, 1999
"... This paper considers the problem of learning the ranking of ' a set of ' stochastic alternatives based upon incomplete information (i.e., a limited number of samples). We describe a system that, at each decision cycle, outputs either a complete ordering on the hypotheses or decides to gather additio ..."
Abstract

Cited by 4 (2 self)
 Add to MetaCart
This paper considers the problem of learning the ranking of ' a set of ' stochastic alternatives based upon incomplete information (i.e., a limited number of samples). We describe a system that, at each decision cycle, outputs either a complete ordering on the hypotheses or decides to gather additional information (Le., observations) at some cost. The ranking problem is a generalization of the previously studied hypothesis selection problem in selection, an algorithm must select the single best hypothesis, while in ranking, an algorithm must order all the hypotheses. The central problem we address is achieving the desired ranking quality while minimizing the cost of acquiring additional samples. We describe two algorithms for hypothesis ranking and their application for the probably approximately correct (PAC) and expecked loss (EL) learning criteria. Empirical results are provided to demonstrate the effectiveness of these ranking procedures on both synthetic and realworld datasets. 1.
Automated Design of KnowledgeLean Heuristics: Learning, Resource Scheduling, and Generalization
, 1996
"... In this thesis we present new methods for the automated design of new heuristics in knowledgelean applications and for finding heuristics that can be generalized to unlearned test cases. These applications lack domain knowledge for credit assignment; hence, operators for composing new heuristics ar ..."
Abstract

Cited by 2 (1 self)
 Add to MetaCart
In this thesis we present new methods for the automated design of new heuristics in knowledgelean applications and for finding heuristics that can be generalized to unlearned test cases. These applications lack domain knowledge for credit assignment; hence, operators for composing new heuristics are generally model free, domain independent, and syntactic in nature. The operators we have used are genetics based; examples of which include mutation and crossover. Learning is based on a generateandtest paradigm that maintains a pool of competing heuristics, tests them to a limited extent, creates new ones from those that perform well in the past, and prunes poor ones from the pool. We have studied four important issues in learning better heuristics: (a) partitioning of a problem domain into smaller subsets, called subdomains, so that performance values within each subdomain can be evaluated statistically, (b) anomalies in performance evaluation within a subdomain, (c) rational scheduling of limited computational resources in testing candidate heuristics in singleobjective as well as multiobjective learning, and (d) finding heuristics that can be generalized to unlearned subdomains. We show experimental results in learning better heuristics for (a) process placement for distributedmemory multicomputers, (b) node decomposition in a branchandbound search, (c) generation of test patterns in VLSI circuit testing, (d) VLSI cell placement and routing, and (e) blind equalization.
Adapting Control Methods for Autonomous Exploration of Unknown Environments, Jet propulsion
"... Abstract On the other hand, developing and Proposed missions to explore comets and moons will encounter environments that are hostile and unpredictable. Any successful explorer must be able to adapt to a wide range of possible operating conditions domainspecific control methods is extremely difficu ..."
Abstract

Cited by 1 (0 self)
 Add to MetaCart
Abstract On the other hand, developing and Proposed missions to explore comets and moons will encounter environments that are hostile and unpredictable. Any successful explorer must be able to adapt to a wide range of possible operating conditions domainspecific control methods is extremely difficult, and requires support of a domain expert. Moreover, the domain expert must have knowledge about the environment in which the spacein order to survive. The traditional approach of concraft is operating, which is not available before structing specialpurpose control methods would require a information about the environment, which is not available a priori for these missions. An alternate approach is to utilize a general control approach with significant capability to adapt its behavior, a so called the spacecraft arrives at the location to explore. If experts are not available, the spacecraft must be able to automatically adapt a flexible control structure specific to the new environment. uduptive problemsolving methodology. Using adaptive problemsolving, a spacecraft can use reinforcement learning to adapt an environmentspecific search strategy given the craft’s general problem solver with a flexible control architecture. The resulting methods would enable the spacecraft increase its perforAdaptive problem solving addresses these problems by enabling the development and maintenance of effective control strategies without extensive domainspecific knowledge. An adaptive problem solver is given: (1) a generic set of conmance with respect to probability of survival and mistrol strategies and (2) a flexible control architecsion goals. ture, and uses a statistical method to estimate the quality of each control strategy or generate a more
Restart Scheduling for Genetic Algorithms
 Genetic Algorithms: Proceedings of the Seventh International Conference
, 1998
"... In order to escape from local optima, it is standard practice to periodically restart a genetic algorithm according to some restart criteria/policy. This paper addresses the issue of finding a good restart strategy in the context of resourcebounded optimization scenarios, in which the goal is t ..."
Abstract
 Add to MetaCart
In order to escape from local optima, it is standard practice to periodically restart a genetic algorithm according to some restart criteria/policy. This paper addresses the issue of finding a good restart strategy in the context of resourcebounded optimization scenarios, in which the goal is to generate the best possible solution given a fixed amount of time. We propose the use of a restart scheduling strategy which generates a static restart strategies with optimal expected utility, based on a database of past performance of the algorithm on a class of problem instances. We show that the performance of static restart schedules generated by the approach can be competitive to that of a commonly used dynamic restart strategy based on detection of lack of progress.