Machine Learning
http://hdl.handle.net/20.500.11824/12
2022-11-30T15:23:05ZTrajectory optimization of space vehicle in rendezvous proximity operation with evolutionary feasibility conserving techniques
http://hdl.handle.net/20.500.11824/1532
Trajectory optimization of space vehicle in rendezvous proximity operation with evolutionary feasibility conserving techniques
Shirazi, A.; Ceberio, J.; Lozano, J.A.
In this paper, a direct approach is developed for discovering optimal transfer trajectories of close-range rendezvous of satellites considering disturbances in elliptical orbits. The control vector representing the inputs is parameterized via different interpolation methods, and an Estimation of Distribution Algorithm (EDA) that implements mixtures of probability models is presented. To satisfy the terminal conditions, which are represented as non-linear inequality constraints, several feasibility conserving mechanisms associated with learning and sampling methods of the EDAs are proposed, which guarantee the feasibility of the explored solutions. They include a particular implementation of a clustering algorithm, outlier detection, and several heuristic mapping methods. The combination of the proposed operators guides the optimization process in achieving the optimal solution by surfing the regions of the search domain associated with feasible solutions. Numerical simulations confirm that space transfer trajectories with minimum-fuel consumption for the chaser spacecraft can be obtained with terminal condition satisfaction in rendezvous proximity operation.
2022-10-09T00:00:00ZAd-Hoc Explanation for Time Series Classification
http://hdl.handle.net/20.500.11824/1517
Ad-Hoc Explanation for Time Series Classification
Abanda, A.; Mori, U.; Lozano, J.A.
In this work, a perturbation-based model-agnostic explanation method for time series classification is presented. One of the main novelties of the proposed method is that the considered perturbations are interpretable and specific for time series. In real-world time series, variations in the speed or the scale of a particular action, for instance, may determine the class, so modifying this type of characteristic leads to ad-hoc explanations for time series. To this end, four perturbations or transformations are proposed: warp, scale, noise, and slice. Given a transformation, an interval of a series is considered relevant for the prediction of a classifier if a transformation in this interval changes the prediction. Another novelty is that the method provides a two-level explanation: a high-level explanation, where the robustness of the prediction with respect to a particular transformation is measured, and a low-level explanation, where the relevance of each region of the time series in the prediction is visualized. In order to analyze and validate our proposal, first some illustrative examples are provided, and then a thorough quantitative evaluation is carried out using a specifically designed evaluation procedure.
2022-01-01T00:00:00ZLearning a Battery of COVID-19 Mortality Prediction Models by Multi-objective Optimization
http://hdl.handle.net/20.500.11824/1515
Learning a Battery of COVID-19 Mortality Prediction Models by Multi-objective Optimization
Martinez, M.; García-Gutierrez, S.; Armañanzas, R.; Díaz, A.; Inza, I.; Lozano, J.A.
The COVID-19 pandemic is continuously evolving with drastically changing epidemiological situations which are approached with different decisions: from the reduction of fatalities to even the selection of patients with the highest probability of survival in critical clinical situations. Motivated by this, a battery of mortality prediction models with different performances has been developed to assist physicians and hospital managers. Logistic regression, one of the most popular classifiers within the clinical field, has been chosen as the basis for the generation of our models. Whilst a standard logistic regression only learns a single model focusing on improving accuracy, we propose to extend the possibilities of logistic regression by focusing on sensitivity and specificity. Hence, the log-likelihood function, used to calculate the coefficients in the logistic model, is split into two objective functions: one representing the survivors and the other for the deceased class. A multi-objective optimization process is undertaken on both functions in order to find the Pareto set, composed of models not improved by another model in both objective functions simultaneously. The individual optimization of either sensitivity (deceased patients) or specificity (survivors) criteria may be conflicting objectives because the improvement of one can imply the worsening of the other. Nonetheless, this conflict guarantees the output of a battery of diverse prediction models. Furthermore, a specific methodology for the evaluation of the Pareto models is proposed. As a result, a battery of COVID-19 mortality prediction models is obtained to assist physicians in decision-making for specific epidemiological situations.
2022-07-09T00:00:00ZThe role of asymmetric prediction losses in smart charging of electric vehicles
http://hdl.handle.net/20.500.11824/1514
The role of asymmetric prediction losses in smart charging of electric vehicles
Straka, M.; Buzna, L.; Refa, N.; Mazuelas, S.
Climate change prompts humanity to look for decarbonisation opportunities, and a viable option is to supply electric vehicles with renewable energy. The stochastic nature of charging demand and renewable generation requires intelligent charging driven by predictions of charging behaviour. The conventional prediction models of charging behaviour usually minimise the quadratic loss function. Moreover, the adequacy of predictions is almost solely evaluated by accuracy measures, disregarding the consequences of prediction losses in an application context. Here, we study the role of asymmetric prediction losses which enable balancing the over- and under-predictions and adjust predictions to smart charging algorithms. Using the main classes of machine learning methods, we trained prediction models of the connection duration and compared their performance for various asymmetries of the loss function. In addition, we proposed a methodological approach to quantify the consequences of prediction losses on the performance of selected archetypal smart charging schemes. In concrete situations, we demonstrated that an appropriately selected degree of the loss function asymmetry is crucial as it almost doubles the price range where the smart charging is beneficial, and increases the extent to which the charging demand is satisfied up to 40%. Additionally, the proposed methods improve charging fairness since the distribution of unmet charging demand across vehicles becomes more homogeneous.
2022-07-01T00:00:00Z