Insight Centre for Data Analytics - Conference Items
Permanent URI for this collection
Browse
Recent Submissions
Item BHO-MA: Bayesian Hyperparameter Optimization with Multi-objective Acquisition(Springer Nature, 2024-02-01) Dogan, Vedat; Prestwich, Steven; Science Foundation IrelandGood hyperparameter values are crucial for the performance of machine learning models. In particular, poorly chosen values can cause under- or overfitting in regression and classification. A common approach to hyperparameter tuning is grid search, but this is crude and computationally expensive, and the literature contains several more efficient automatic methods such as Bayesian optimization. In this work, we develop a Bayesian hyperparameter optimization technique with more robust performance, by combining several acquisition functions and applying a multi-objective approach. We evaluated our method using both classification and regression tasks. We selected four data sets from the literature and compared the performance with eight popular methods. The results show that the proposed method achieved better results than all others.Item Modeling robustness in CSPs as weighted CSPs(Springer, 2013) Climent, Laura; Wallace, Richard J.; Salido, Miguel A.; Barber, Federico; Ministerio de Ciencia, Tecnología e InnovaciónMany real life problems come from uncertain and dynamic environments, where the initial constraints and/or domains may undergo changes. Thus, a solution found for the problem may become invalid later. Hence, searching for robust solutions for Constraint Satisfaction Problems (CSPs) becomes an important goal. In some cases, no knowledge about the uncertain and dynamic environment exits or it is hard to obtain it. In this paper, we consider CSPs with discrete and ordered domains where only limited assumptions are made commensurate with the structure of these problems. In this context, we model a CSP as a weighted CSP (WCSP) by assigning weights to each valid constraint tuple based on its distance from the edge of the space of valid tuples. This distance is estimated by a new concept introduced in this paper: coverings. Thus, the best solution for the modeled WCSP can be considered as a robust solution for the original CSP according to our assumptions.Item Extrapolating from limited uncertain information to obtain robust solutions for large-scale optimization problems(Institute of Electrical and Electronics Engineers (IEEE), 2014-12-15) Climent, Laura; Wallace, Richard; O'Sullivan, Barry; Freuder, Eugene; Science Foundation IrelandData uncertainty in real-life problems is a current challenge in many areas, including Operations Research (OR) and Constraint Programming (CP). This is especially true given the continual and accelerating increase in the amount of data associated with real-life problems, to which Large Scale Combinatorial Optimization (LSCO) techniques may be applied. Although data uncertainty has been studied extensively in the literature, many approaches do not take into account the partial or complete lack of information about uncertainty in real-life settings. To meet this challenge, in this paper we present a strategy for extrapolating data from limited uncertain information to ensure a certain level of robustness in the solutions obtained. Our approach is motivated by real-world applications of supply of timber from forests to saw-mills.Item A fully Bayesian approach to bilevel problems(Springer Nature, 2024-10-16) Dogan, Vedat; Prestwich, Steven; O’Sullivan, Barry; Science Foundation IrelandThe mathematical models of many real-world decision-making problems contain two levels of optimization. In these models, one of the optimization problems appears as a constraint of the other one, called follower and leader, respectively. These problems are known as bilevel optimization problems (BOPs) in mathematical programming and are widely studied by both classical and evolutionary optimization communities. The nested nature of these problems causes many difficulties such as non-convexity and disconnectedness for traditional methods, and requires a huge number of function evaluations for evolutionary algorithms. This paper proposes a fully Bayesian optimization approach, called FB-BLO. We aim to reduce the necessary function evaluations for both upper and lower level problems by iteratively approximating promising solutions with Gaussian process surrogate models at both levels. The proposed FB-BLO algorithm uses the other decision-makers’ observations in its Gaussian process model to leverage the correlation between decisions and objective values. This allows us to extract knowledge from previous decisions for each level. The algorithm has been evaluated on numerous benchmark problems and compared with existing state-of-the-art algorithms. Our evaluation demonstrates the success of our proposed FB-BLO algorithm in terms of both effectiveness and efficiency.Item A hybrid Bayesian approach for pessimistic bilevel problems with a new formulation(2024) Dogan, Vedat; Prestwich, Steven D.; O'Sullivan, Barry; Science Foundation IrelandIn many real-world problems, finding the optimal decision for a decision-maker depends on another decision-maker’s response, and it is called bilevel optimization in mathematical programming. It contains two levels of optimization problems while one appears as a constraint of another one called follower and leader, respectively. In many real-world scenarios, the lower level has multiple global optima and the upper level needs to make worst-case assumptions about the decision of the lower level, called the pessimistic case of the bilevel problem. Various approaches have been implemented over the years to solve generic bilevel problems, but few of them could be extended to pessimistic cases. In this short paper, we first propose a new formulation for the pessimistic case. In this way, we take advantage of the hierarchical structure of bilevel problems to make the results more accurate for pessimistic cases. Then, we implement a black-box approach to solve the pessimistic upper level problem to decrease the necessary function evaluations. The performance of the problem is examined by solving a test benchmark problem from the literature.