Multi-method Ensemble Research Articles

Evolving interpretable decision trees for reinforcement learning

In recent years, reinforcement learning (RL) techniques have achieved great success in many different applications. However, their heavy reliance on complex deep neural networks makes most RL models uninterpretable, limiting their application in domains where trust and security are important. To address this challenge, we propose MENS-DT-RL, an algorithm capable of constructing interpretable models for RL via the evolution of decision tree (DT) models. MENS-DT-RL uses a multi-method ensemble algorithm to evolve univariate DTs, guiding the process with a fitness metric that prioritizes interpretability and consistent high performance. Three different initializations for the MENS-DT-RL are proposed, including the use of Imitation Learning (IL) techniques, and a novel pruning approach that reduces solution size without compromising performance. To evaluate the proposed approach, we compare it with other models from the literature on three benchmark tasks from the OpenAI Gym library, as well as on a fertilization problem inspired by real-world crop management. To the best of our knowledge, the proposed scheme is the first to solve the Lunar Lander benchmark with both interpretability and a high confidence rate (90% of episodes are successful), as well as the first to solve the Mountain Car environment with a tree of only 7 nodes. On the real-world task, the proposed MENS-DT-RL is able to produce solutions with the same quality as deep RL policies, with the added bonus of interpretability. We also analyze the best solutions found by the algorithm and show that they are not only interpretable but also diverse in their behavior, empowering the end user with the choice of which model to apply. Overall, the findings show that the proposed approach is capable of producing high-quality transparent models for RL, achieving interpretability without losing performance.

Projected heat wave increasing trends over China based on combined dynamical and multiple statistical downscaling methods

Extensive investigations on the projection of heat waves (HWs) were conducted on the basis of coarse-resolution global climate models (GCMs). However, these investigations still fail to characterise the future changes in HWs regionally over China. PRECIS dynamical downscaling with a horizontal resolution of 25 km × 25 km was employed on the basis of GCM-HadCM3 to provide reliable projections on HWs over the Chinese mainland, and six statistical downscaling methods were used for bias correction under RCP4.5 and RCP8.5 scenarios. The multi-method ensemble (MME) of the top three dynamical downscaling methods with good performance was used to project future changes. Results showed that PRECIS primarily replicated the detailed spatiotemporal pattern of HWs. However, PRECIS overestimated the HWs in the Northwest and Southeast and expanded the areas of HWs in the Northeast and Southwest. Three statistical downscaling methods (quantile mapping, CDF-t and quantile delta mapping) demonstrated good performance in improving PRECIS simulation for reproducing HWs. By contrast, parametric-based trend-preserving approaches such as scaled distribution mapping and ISI-MIP are outperformed by the three aforementioned methods in downscaling HWs, particularly in the high latitudes of China. Based on MME projections, at the end of the 21st century, the national average of the number of HW days each year, the length of the longest HW event in the year and the extreme maximum temperature in HW will increase by 3 times, 1 time and 1.3 °C, respectively, under the RCP4.5 scenario, whilst that under the RCP8.5 scenario will increase by 8 times, 3 times and 3.7 °C, respectively, relative to 1986–2005. The Northwest is regionally projected to suffer long and hot HWs, whilst the South and Southeast will experience frequent consecutive HWs. Thus, HWs projected by the combined dynamical and statistical downscaling method are highly reliable in projecting HWs over China.

A method review of the climate change impact on crop yield

Climate change significantly impacts global agricultural production, giving rise to considerable uncertainties. To explore these climate impacts, three independent methods have been employed: manipulated experiments, process-based crop models, and empirical statistical models. However, the uncertainty stemming from the use of different methods has received insufficient attention, and its implications remain unclear, necessitating a systematic review. In this study, we conducted a comprehensive review of numerous previous studies to summarize the historic development and current status of each method. Through a method comparison, we identified their respective strengths, limitations, and ideal areas of application. Additionally, we outlined potential prospects and suggested directions for future improvements, including clarifying the response mechanisms, updating simulation technologies, and developing multi-method ensembles. By addressing the knowledge gap regarding method differences, this review could contribute to a more accurate assessment of climate impacts on agriculture.

New Probabilistic, Dynamic Multi-Method Ensembles for Optimization Based on the CRO-SL

In this paper, new probabilistic and dynamic (adaptive) strategies for creating multi-method ensembles based on the coral reef optimization with substrate layers (CRO-SL) algorithm are proposed. CRO-SL is an evolutionary-based ensemble approach that is able to combine different search procedures for a single population. In this work, two different probabilistic strategies to improve the algorithm are analyzed. First, the probabilistic CRO-SL (PCRO-SL) is presented, which substitutes the substrates in the CRO-SL population with tags associated with each individual. Each tag represents a different operator which will modify the individual in the reproduction phase. In each generation of the algorithm, the tags are randomly assigned to the individuals with similar probabilities, obtaining this way an ensemble that sees more intense changes with the application of different operators to a given individual than CRO-SL. Second, the dynamic probabilistic CRO-SL (DPCRO-SL) is presented, in which the probability of tag assignment is modified during the evolution of the algorithm, depending on the quality of the solutions generated in each substrate. Thus, the best substrates in the search process will be assigned higher probabilities than those which showed worse performance during the search. The performances of the proposed probabilistic and dynamic ensembles were tested for different optimization problems, including benchmark functions and a real application of wind-turbine-layout optimization, comparing the results obtained with those of existing algorithms in the literature.

Assessment of future wind resources under climate change using a multi-model and multi-method ensemble approach

Harvesting renewable wind energy is among the most cost-efficient means of reducing carbon emissions and achieving carbon neutrality. However, wind resource is susceptible to climate change impacts since the global temperature increase will reshape atmospheric circulation patterns. To facilitate the evaluation of fine-scale wind energy potential under climate variability, this paper proposes and validates a multi-model and multi-method ensemble wind resource projection approach. Then this approach is utilized to investigate the future variation of wind resources in Hong Kong based on the combined use of global climate models from Coupled Model Intercomparison Project Phase 6 and long-term observations from meteorological stations. It is found that there is a significant increase in future wind resources during summer, while a remarkable decline is projected during autumn. Nevertheless, the variations of wind resources in winter and spring are relatively insignificant. The outcomes of this study are expected to offer a framework for fine-scale wind resource assessment under climate change, and facilitate the economic and risk assessments of future wind farm projects.

A versatile multi-method ensemble for wind farm layout optimization

In this paper, the performance of a multi-method ensemble for optimization in a turbine layout problem for wind farms is discussed. Specifically, the Coral Reefs Optimization algorithm with Substrate Layer (CRO-SL) is evaluated. The CRO-SL is an evolutionary-type meta-heuristic, able to combine different search procedures or methods within a single population, leading to a multi-method ensemble. The CRO-SL has been adapted to the wind turbines layout problem by including a discrete encoding, based on sparse matrices (though it can also work over continuous encodings) and a convex geometric procedure to speed-up the wake effect calculation for each turbine. The performance of the proposed CRO-SL algorithm in the turbine layout problem is evaluated by means of different experiments, considering simulated and real data from a wind farm in Spain. The performance of the CRO-SL is also evaluated in a different wind turbine layout problem from a NREL-IEA case competition. The contribution of each substrate of the CRO-SL algorithm (search procedure) is discussed in terms of the global search ability of the algorithm in the different problems considered. The global performance of the proposed approach has been finally compared to that of existing approaches in the literature, obtaining excellent results in all scenarios considered.

Investigation and evaluation of algorithms for unmanned aerial vehicle multispectral image registration

Automatically registration of unmanned aerial vehicle (UAV) multispectral images is fundamental for subsequent applications. Although many studies exist for visible camera images and satellite multispectral image registration, studies for UAV are still rare. Under this context, this study firstly evaluates the performance of several widely used traditional methods (i.e., SIFT, SURF, ORB, and CFOG) for visible camera and satellite images in UAV multispectral image registration. This study further proposes an unsupervised and end-to-end deep learning network (i.e., DSIM) for multispectral image registration. An evident feature of DSIM is to regress the homography parameters with convolutional neural networks and to use the pyramid structural similarity loss to optimize the network. 1200 groups of UAV multispectral images acquired over three different sites in four months are used to comprehensively test the aforementioned five methods. Results show that CFOG achieves the highest correct matching rate in the test set, followed by DSIM and SIFT. Nevertheless, DSIM is more robust in images with weak or repeated texture than CFOG and SIFT. In addition, performance of CFOG and SIFT is closely related to the number of the found matching points. Based on the findings, we propose a multi-method ensemble strategy to combine CFOG, DSIM, and SIFT according to the number of the found matching points. This strategy outperforms the individual methods with a correct matching rate of 96.2%. Lower correct matching rate of CFOG + SIFT confirms that DSIM and traditional methods are very complementary in UAV multispectral image registrations.

Holocene global mean surface temperature, a multi-method reconstruction approach

An extensive new multi-proxy database of paleo-temperature time series (Temperature 12k) enables a more robust analysis of global mean surface temperature (GMST) and associated uncertainties than was previously available. We applied five different statistical methods to reconstruct the GMST of the past 12,000 years (Holocene). Each method used different approaches to averaging the globally distributed time series and to characterizing various sources of uncertainty, including proxy temperature, chronology and methodological choices. The results were aggregated to generate a multi-method ensemble of plausible GMST and latitudinal-zone temperature reconstructions with a realistic range of uncertainties. The warmest 200-year-long interval took place around 6500 years ago when GMST was 0.7 °C (0.3, 1.8) warmer than the 19th Century (median, 5th, 95th percentiles). Following the Holocene global thermal maximum, GMST cooled at an average rate −0.08 °C per 1000 years (−0.24, −0.05). The multi-method ensembles and the code used to generate them highlight the utility of the Temperature 12k database, and they are now available for future use by studies aimed at understanding Holocene evolution of the Earth system.

Addressing Unequal Area Facility Layout Problems with the Coral Reef Optimization algorithm with Substrate Layers

The Unequal Area Facility Layout Problem (UA-FLP) is a relevant task in industrial manufacturing, in which the disposition of a number of facilities (or departments) in a manufacturing system must be obtained, under several optimization criteria and different constraints. The UA-FLP is a hard optimization problem, in which traditional optimization techniques do not obtain good results. Thus, it has been successfully tackled with different heuristics and meta-heuristics in the last years. In this work we address the UA-FLP with a multi-method ensemble approach, the Coral Reefs Optimization algorithm with Substrate Layers (CRO-SL). It is a novel multi-method evolutionary algorithm that encourages the evolution of several searching procedures at the same time over a single population. The CRO-SL has been previously applied to very difficult optimization problems, obtaining excellent performance. In this case, we adapt the CRO-SL to the UA-FLP, by means of increasing the diversity generation within the algorithm, which is helpful to improve the exploration of the searching space, avoiding to fall into local minima. Specifically, we propose to include several reproduction mechanisms (adapted to the UA-FLP) within each substrate of the algorithm, which will highly increase the diversity generation in the CRO-SL. An exhaustive experimental study of the CRO-SL performance in a large number of UA-FLP instances is carried out, including a comparison with the state-of-the-art algorithms for this problem. We will show the ability of the CRO-SL to reach or surpass the best-known solutions in most of the tested UA-FLP cases.

Producing actionable climate change information for regions: the distillation paradigm and the 3R framework

The issue of the distillation of actionable climate change information for application to vulnerability, impacts and adaptation studies in support of climate service activities (or VIA-CS) at regional to local scales is revisited. The 3R approach is introduced: robustness, reliability and relevance. Climate information for regions needs to be robust in the sense of producing statistically significant and consistent change signals based on multimodel and multimethod ensembles; it needs to be reliable in the sense of being based on a good understanding of the physical processes underlying the change signals, and on models capable of reproducing the functioning of the climate system at different scales; it needs to be relevant in the sense of providing information targeted for use in VIA-CS applications, including a proper characterization of uncertainties based on probabilistic approaches. It is advocated that the distillation problem requires an interdisciplinary consensus approach and a new generation of scientists acting at the interface between the climate modelling and end-user communities.

Robustness and uncertainties in global multivariate wind-wave climate projections

Understanding climate-driven impacts on the multivariate global wind-wave climate is paramount to effective offshore/coastal climate adaptation planning. However, the use of single-method ensembles and variations arising from different methodologies has resulted in unquantified uncertainty amongst existing global wave climate projections. Here, assessing the first coherent, community-driven, multi-method ensemble of global wave climate projections, we demonstrate widespread ocean regions with robust changes in annual mean significant wave height and mean wave period of 5–15% and shifts in mean wave direction of 5–15°, under a high-emission scenario. Approximately 50% of the world’s coastline is at risk from wave climate change, with ~40% revealing robust changes in at least two variables. Furthermore, we find that uncertainty in current projections is dominated by climate model-driven uncertainty, and that single-method modelling studies are unable to capture up to ~50% of the total associated uncertainty.

Building Up a Robust Risk Mathematical Platform to Predict Colorectal Cancer

Colorectal cancer (CRC), as a result of a multistep process and under multiple factors, is one of the most common life-threatening cancers worldwide. To identify the “high risk” populations is critical for early diagnosis and improvement of overall survival rate. Of the complicated genetic and environmental factors, which group is mostly concerning colorectal carcinogenesis remains contentious. For this reason, this study collects relatively complete information of genetic variations and environmental exposure for both CRC patients and cancer-free controls; a multimethod ensemble model for CRC-risk prediction is developed by employing such big data to train and test the model. Our results demonstrate that (1) the explored genetic and environmental biomarkers are validated to connect to the CRC by biological function- or population-based evidences, (2) the model can efficiently predict the risk of CRC after parameter optimization by the big CRC-related data, and (3) our innovated heterogeneous ensemble learning model (HELM) and generalized kernel recursive maximum correntropy (GKRMC) algorithm have high prediction power. Finally, we discuss why the HELM and GKRMC can outperform the classical regression algorithms and related subjects for future study.

Similar estimates of temperature impacts on global wheat yield by three independent methods

The potential impact of global temperature change on global crop yield has recently been assessed with different methods. Here we show that grid-based and point-based simulations and statistical regressions (from historic records), without deliberate adaptation or CO 2 fertilization effects, produce similar estimates of temperature impact on wheat yields at global and national scales. With a 1 °C global temperature increase, global wheat yield is projected to decline between 4.1% and 6.4%. Projected relative temperature impacts from different methods were similar for major wheat-producing countries China, India, USA and France, but less so for Russia. Point-based and grid-based simulations, and to some extent the statistical regressions, were consistent in projecting that warmer regions are likely to suffer more yield loss with increasing temperature than cooler regions. By forming a multi-method ensemble, it was possible to quantify 'method uncertainty' in addition to model uncertainty. This significantly improves confidence in estimates of climate impacts on global food security.

Multi-method ensemble selection of spectral bands related to leaf biochemistry

Multi-method ensembles are generally believed to return more reliable results than the application of one method alone. Here, we test if for the quantification of leaf traits an ensemble of regression models, consisting of Partial Least Squares (PLSR), Random Forest (RFR), and Support Vector Machine regression (SVMR) models, is able to improve the robustness of the spectral band selection process compared to the outcome of a single technique alone. The ensemble approach was tested using one artificial and five measured data sets of leaf level spectra and corresponding information on leaf chlorophyll, dry matter, and water content. PLSR models optimized for the goodness of fit, an established approach for band selection, were used to evaluate the performance of the ensemble. Although the fits of the models within the ensemble were poorer than the fits achieved with the reference approach, the ensemble was able to provide a band selection with higher consistency across all data sets. Due to the selection characteristics of the methods within the ensemble, the ensemble selection is moderately narrow and restrictive but in good agreement with known absorption features published in literature. We conclude that analyzing the range of agreement of different model types is an efficient way to select a robust set of spectral bands related to the foliar properties under investigation. This may help to deepen our understanding of the spectral response of biochemical and biophysical traits in foliage and canopies.

Mapping seabed sediments: Comparison of manual, geostatistical, object-based image analysis and machine learning approaches

Marine spatial planning and conservation need underpinning with sufficiently detailed and accurate seabed substrate and habitat maps. Although multibeam echosounders enable us to map the seabed with high resolution and spatial accuracy, there is still a lack of fit-for-purpose seabed maps. This is due to the high costs involved in carrying out systematic seabed mapping programmes and the fact that the development of validated, repeatable, quantitative and objective methods of swath acoustic data interpretation is still in its infancy. We compared a wide spectrum of approaches including manual interpretation, geostatistics, object-based image analysis and machine-learning to gain further insights into the accuracy and comparability of acoustic data interpretation approaches based on multibeam echosounder data (bathymetry, backscatter and derivatives) and seabed samples with the aim to derive seabed substrate maps. Sample data were split into a training and validation data set to allow us to carry out an accuracy assessment. Overall thematic classification accuracy ranged from 67% to 76% and Cohen׳s kappa varied between 0.34 and 0.52. However, these differences were not statistically significant at the 5% level. Misclassifications were mainly associated with uncommon classes, which were rarely sampled. Map outputs were between 68% and 87% identical. To improve classification accuracy in seabed mapping, we suggest that more studies on the effects of factors affecting the classification performance as well as comparative studies testing the performance of different approaches need to be carried out with a view to developing guidelines for selecting an appropriate method for a given dataset. In the meantime, classification accuracy might be improved by combining different techniques to hybrid approaches and multi-method ensembles.

Dynamic-Model-Based Seasonal Prediction of Meteorological Drought over the Contiguous United States

Abstract A simple method was developed to forecast 3- and 6-month standardized precipitation indices (SPIs) for the prediction of meteorological drought over the contiguous United States based on precipitation seasonal forecasts from the NCEP Climate Forecast System (CFS). Before predicting SPI, the precipitation (P) forecasts from the coarse-resolution CFS global model were bias corrected and downscaled to a regional grid of 50 km. The downscaled CFS P forecasts, out to 9 months, were appended to the P analyses to form an extended P dataset. The SPIs were calculated from this new time series. Five downscaling methods were tested: 1) bilinear interpolation; 2) a bias correction and spatial downscaling (BCSD) method based on the probability distribution functions; 3) a conditional probability estimation approach using the mean P ensemble forecasts developed by J. Schaake, 4) a Bayesian approach that bias corrects and downscales P using all ensemble forecast members, as developed by the Princeton University group; and 5) multimethod ensemble as the equally weighted mean of the BCSD, Schaake, and Bayesian forecasts. For initial conditions from April to May, statistical downscaling methods were compared with dynamic downscaling based on the NCEP regional spectral model and forecasts from a high-resolution CFS T382 model. The skill is regionally and seasonally dependent. Overall, the 6-month SPI is skillful out to 3–4 months. For the first 3-month lead times, forecast skill comes from the P analyses prior to the forecast time. After 3 months, the multimethod ensemble has small advantages, but forecast skill may be too low to be useful in practice.

Machine Learning Methods Applied to Pharmacokinetic Modelling of Remifentanil in Healthy Volunteers: A Multi-Method Comparison

This study compared the blood concentrations of remifentanil obtained in a previous clinical investigation with the predicted remifentanil concentrations produced by different pharmacokinetic models: a non-linear mixed effects model created by the software NONMEM; an artificial neural network (ANN) model; a support vector machine (SVM) model; and multi-method ensembles. The ensemble created from the mean of the ANN and the non-linear mixed effects model predictions achieved the smallest error and the highest correlation coefficient. The SVM model produced the highest error and the lowest correlation coefficient. Paired t-tests indicated that there was insufficient evidence that the predicted values of the ANN, SVM and two multi-method ensembles differed from the actual measured values at alpha = 0.05. The ensemble method combining the ANN and non-linear mixed effects model predictions outperformed either method alone. These results indicated a potential advantage of ensembles in improving the accuracy and reducing the variance of pharmacokinetic models.

Population Pharmacokinetic and Pharmacodynamic Models of Propofol in Healthy Volunteers using NONMEM and Machine Learning Methods

Objectives: The primary objective of this study is to compare model performance of machine learning methods with that of a previous study in which a nonlinear mixed effects model was created using NONMEM? for the pharmacokinetic and pharmacodynamic data for propofol. The secondary objective was to evaluate if a pharmacodynamic model describing the relationship between the dose of propofol and bispectral index (BIS) outperform that describing the relationship between a pharmacokinetic model derived-predicted concentrations of propofol and BIS. Methods: Data were collected during a study involving the infusion of propofol into healthy volunteers. Pharmacokinetic and pharmacodynamic models were constructed using artificial neural networks (ANNs), support vector machines (SVMs), and multi-method ensembles and were compared with the nonlinear mixed effects method as implemented by NONMEM?. Model performance was assessed by goodnessoffit statistics, paired t-tests between predicted and observed values for each model and scatterplots. Results: In pharmacokinetic analysis, ensemble I, the mean of ANN and NONMEM? predictions, achieved minimal error and the highest correlation coefficient. SVM produced the highest error and the lowest correlation coefficient. In pharmacodynamic analysis, ANN exhibited the best performance. An ANN Model describing the relationship between the dose of propofol and BIS was not inferior to an ANN model describing the relationship between predicted concentrations of propofol derived from an ANN pharmacokinetic model and BIS. Conclusions: In pharmacokinetic analysis, ensemble combined with ANN achieved slightly better performance than NONMEM?. The relationship between the dose of propofol and BIS can be predicted without considering pharmacokinetics of propofol. (Journal of Korean Society of Medical Informatics 14-2, 147-159, 2008)