Over the years, researchers have proposed multiple approaches to reduce the number of parameters Deep Learning models have. Due to the complexity of compressing models, some authors have opted to ...train Reinforcement Learning agents that learn how to compress a particular model without losing considerable accuracy. Nonetheless, training an agent for each model can be time-consuming. We propose a methodology for training a generalist agent capable of compressing other convolutional neural networks that it was not trained to compress. Our generalist agent uses feature maps to select which compression technique to apply to convolutional and dense layers. Since the shape of the feature maps is reduced as it goes deeper into the network, we implemented a Dueling Deep Q-Network with a Region of Interest layer, allowing it to generate features of a fixed size for feature maps of various heights and widths. Our generalist agent trained to compress two LeNet models, one trained with fashion MNIST and the other with Kuzushiji-MNIST, compressed the same architecture trained on MNIST to less than 15% of its original size with an accuracy loss of less than 2.5%.
The Robotic Mobile Fulfillment System (RMFS) is a method for handling products, in which a Line Follower Robot (LFR) transports products to a human workstation for packing. In this systematic review, ...we delve into the current state of RMFS research using data sourced from Scopus. After a comprehensive search, we found 264 manuscripts, which we filtered to 76 relevant articles. Our analysis covers several variables, from basic metadata to manuscript impact and specific conditions the authors consider. We discovered that there needs to be more focus on the pod allocation problem, despite its potential, with the majority of the emphasis on LFR displacement. We created a detailed diagram that outlines the essential elements and subproblems associated with RMFS. As the interest in RMFS continues growing, our study provides crucial insights and direction for future research efforts.
Much research aims to enhance weapon detection by applying different techniques to object detection models. However, little research focuses on identifying armed people through real-time surveillance ...cameras. The proposed solution involves the development of algorithms for identifying people carrying handguns (pistols and revolvers). We have chosen the YOLOv4 model to detect people, guns, and faces. Then, we extract information from YOLO related to real-time videos, such as bounding box coordinates, distances, and intersection areas between firearms and the people in each video frame to recognize the armed people. There are some challenges to overcome, for example, occlusion, hidden handguns, and people close to each other. It allows us to develop and compare different types of solutions. We proposed three heuristics and seven machine-learning models. The heuristics are the method of centers, the method of intersections, and the method of distances. Furthermore, the machine learning models are Random Forest Classifier, Multilayer Perceptron, k-Nearest-Neighbors, Support Vector Machine, Logistic Regression, Naive Bayes, and Gradient Boosting Classifier. The Random Forest Classifier presented the best performance reaching an accuracy of 85.44%, a precision of 87.07%, a recall of 88.68%, and an F1-score of 87.87%.
Hyper-heuristics aim at interchanging different solvers while solving a problem. The idea is to determine the best approach for solving a problem at its current state. This way, every time we make a ...move it gets us closer to a solution. The problem changes; so does its state. As a consequence, for the next move, a different solver may be invoked. Hyper-heuristics have been around for almost 20 years. However, combinatorial optimization problems date from way back. Thus, it is paramount to determine whether the efforts revolving around hyper-heuristic research have been targeted at the problems of the highest interest for the combinatorial optimization community. In this work, we tackle such an endeavor. We begin by determining the most relevant combinatorial optimization problems, and then we analyze them in the context of hyper-heuristics. The idea is to verify whether they remain as relevant when considering exclusively works related to hyper-heuristics. We find that some of the most relevant problem domains have also been popular for hyper-heuristics research. Alas, others have not and few efforts have been directed towards solving them. We identify the following problem domains, which may help in furthering the impact of hyper-heuristics: Shortest Path, Set Cover, Longest Path, and Minimum Spanning Tree. We believe that focusing research on ways for solving them may lead to an increase in the relevance and impact that hyper-heuristics have on combinatorial optimization problems.
Hyper-heuristics (HHs) stand as a relatively recent approach to solving optimization problems. There are different kinds of HHs. One of them deals with how low-level heuristics must be combined to ...deliver an improved solution to a set of problem instances. Literature commonly refers to them as selection hyper-heuristics . One of their advantages is that the strengths of each heuristic can be fused into a high-level solver. However, one of their drawbacks is that sometimes this generalization scheme does not suffice. Additionally, it is not easy to reuse these HHs since the model cannot be easily tweaked. So, in this work, we develop a hyper-heuristic model with an additional layer of generalization. The rationale behind it is to preserve the general structure of selecting an adequate solver for a particular situation but to use HHs instead of low-level heuristics. We call this model a Squared Hyper-Heuristic (SHH). To validate our proposal, we pursue a four-stage methodology that covers several testing scenarios. Our data reveal that, under proper conditions, our model can outperform the base HHs. Moreover, it is flexible enough to allow for an increased number of layers so that the complexity of the final model can be tuned. Additionally, different kinds of instances can be used to train each stage of the model, thus setting the groundwork for developing a transfer learning approach for hyper-heuristics.
Selection hyper-heuristics are a technology for optimization in which a high-level mechanism controls low-level heuristics, so as to be capable of solving a wide range of problem instances ...efficiently. Hyper-heuristics are used to generate a solution process rather than producing an immediate solution to a given problem. This process is a re-usable mechanism that can be applied both to seen and unseen problem instances. In this paper, we propose a selection hyper-heuristic process with the intention to rise the level of generality and solve consistently well a wide range of constraint satisfaction problems. The hyper-heuristic technique is based on a messy genetic algorithm that generates high-level heuristics formed by rules (condition
→
heuristic). The high-level heuristics produced are seen to be good at solving instances from certain parts of the parameterized space of problems, producing results using effort comparable to the best single heuristic per instance. This is beneficial, as the choice of best heuristic varies from instance to instance, so the high-level heuristics are definitely preferable to selecting any one low-level heuristic for all instances. The results confirm the robustness of the proposed approach and how high-level heuristics trained for some specific classes of instances can also be applied to unseen classes without significant lost of efficiency. This paper contributes to the understanding of heuristics and the way they can be used in a collaborative way to benefit from their combined strengths.
The Monterrey Metropolitan Area is one of the most densely populated and polluted regions in Latin America. Hence, providing early warnings to the population when pollutant concentrations reach high ...levels is critical. This allows people at higher health risk to make informed decisions about when to go out, mitigating future health complications. Using forecasting models, we can produce timely warnings for future concentration levels. In this work, we implement a set of short-term shallow machine learning models that would serve as a baseline for future forecasting analyses of PM2.5 concentration levels in the Monterrey Metropolitan Area. The proposed approach starts with multiple imputation through chained equations for missing value imputation, the incorporation of time metadata, and target winsorization. Then, we rely on the well-known random search for parameter optimization of the machine learning models and k-fold cross-validation, obtaining favorable results. We devise these models for a single-step and single-station analysis on an hourly multivariate air quality dataset (containing 77203 rows and 16 columns from the first hour of January 1, 2015 00:00:00 to April 17, 2022 23:00:00) and compare them using standard regression metrics. Therefore, we identify the forecasting model with the best performance, which was an Extra Trees Regressor with a Root Mean Squared Error of 0.013, a Mean Absolute Error of 0.006 (equivalent to a Mean Absolute Percentage Error of 0.294% and a Symmetric Mean Absolute Percentage Error of 0.078%), and a Maximum Error of 0.187μg/m3.
Display omitted
•We study various shallow ML algorithms to forecast PM2.5 levels in the MMA.•We provide a working framework in Python to forecast PM2.5 levels in the MMA.•We observed that shallow ML models compete with DL algorithms in forecasting tasks.
In the algorithm selection problem, where the task is to identify the most suitable solving technique for a particular situation, most methods used as performance mapping mechanisms have been ...relatively simple models such as logistic regression or neural networks. In the latter case, most implementations tend to have a shallow and straightforward architecture and, thus, exhibit a limited ability to extract relevant patterns. This research explores the use of attention-based neural networks as meta-learners to improve the performance mapping mechanism in the algorithm selection problem and fully take advantage of the model’s capabilities for pattern extraction. We compare the proposed use of an attention-based meta-learner method as a performance mapping mechanism against five models from the literature: multi-layer perceptron, k-nearest neighbors, softmax regression, support vector machines, and decision trees. We used a meta-data dataset obtained by solving the vehicle routing problem with time window (VRPTW) instances contained in the Solomon benchmark with three different configurations of the simulated annealing meta-heuristic for testing purposes. Overall, the attention-based meta-learner model yields better results when compared to the other benchmark methods in consistently selecting the algorithm that best solves a given VRPTW instance. Moreover, by significantly outperforming the multi-layer perceptron, our findings suggest promising potential in exploring more recent and novel advancements in neural network architectures.
Renewable energy sources are an environmentally attractive idea, but they require a proper control scheme to guarantee optimal operation. In this work, we tune different controllers for an ...Interleaved Boost Converter (IBC) powered by a photovoltaic array using three metaheuristics: Genetic Algorithm, Particle Swarm Optimization, and Gray Wolf Optimization. We also develop several controllers for a second simulated scenario where the IBC is plugged into an existing microgrid (MG) as this can provide relevant data for real-life applications. In both cases, we consider hybrid controllers based on a Linear Quadratic Regulator (LQR). However, we hybridize it with an Integral action (I-LQR) in the first scenario to compare our data against previously published controllers. In the second one, we add a Proportional-Integral technique (PI-LQR) as we do not have previous data to compare against to provide a more robust controller than I-LQR. To validate our approach, we run extensive simulations with each metaheuristic and compare the resulting data. We focus on two fronts: the performance of the controllers and the computing cost of the solvers when facing practical issues. Our results demonstrate that the approach proposed for tuning controllers is a feasible strategy. The controllers tuned with the metaheuristics outperformed previously proposed strategies, yielding solutions thrice faster with virtually no overshoot and a voltage ripple seven times smaller. Not only this, but our controllers could correct some issues liaised to the IBC when it is plugged into an MG. We are confident that these insights can help migrate this approach to a more diverse set of MGs with different renewable sources and escalate it to real-life experiments.