Neural Architecture Search is an active research field that aims to design neural networks automatically. Nevertheless, this is usually an expensive process since the search algorithm must evaluate the performance of many candidate solutions from a vast search space. Because of that, different strategies have been proposed to perform efficient Neural Architecture Search. The recent development of zero-cost performance predictors has shown a lot of promise due to the possibility of predicting a network's performance without training. On the other hand, a predictor's correlation with a model's performance may depend on the network search space and on the benchmark dataset. Each performance predictor might lead the search processes to favor very different network patterns. A design principle is defined as a restriction in a hyperparameter distribution that is expected to yield optimum network performance. In this work, we propose an automatic iterative approach to uncover the design principles of deep neural networks optimized by zero-cost performance predictors, and we discuss insightful information obtained by its application.

Neural Architecture Search (NAS) is a busy research field growing exponentially in recent years. State-of-the-art deep neural networks usually require a specialist to fine-tune the model to solve a specific problem. NAS research aims to design neural network architectures automatically, thus easing the need for machine learning specialists to spend a lot of effort on hand-crafted attempts. As artificial intelligence applications are becoming ubiquitous, there is also a growing interest in efficient applications that could be deployed to smartphones, smart wearable devices, and other edge devices. Gender recognition in unfiltered images — such as those we find in real-world situations like pictures taken with smartphones and video shots from surveillance cameras — is one of such challenging applications. In this work, we developed an evolutionary NAS algorithm that consistently finds efficient ResNet-based architectures, named RENNAS, which have a good trade-off between classification accuracy and architectural and computational complexities. We demonstrate our algorithm's performance on Adience dataset of unfiltered images for gender recognition.

In the context of software development, managing and organizing agile boards of multi-disciplinary teams distributed around the world is a great challenge, especially regarding the process of assigning tickets to the correct team roles. Incorrectly assigned tickets can result in significant resource waste in any project and directly influence delivery outcomes and project costs. This work proposes a method for ticket analysis and automatic team assignment using Natural Language Processing and explainable Machine Learning models. Results show that the models perform well on a real-world team assignment task and provide insights into their decision.

This work presents an analysis of sampled Local Optima Networks (LONs) on flowshop instances. LON metrics are addressed here to scrutinize the performance of four sampling strategies, each one resulting from the combination of a local search (iterative or best improvement) and a perturbation operator (deconstruction with or without local search on partial solutions). The results highlight the superiority of iterative improvement and the advantages of exploiting partial solutions, especially when they are combined to discover new local optima regions on large problems. LON metrics are also useful for predicting performance of optimization algorithms - in particular, this work considers two Iterated Greedy variants when solving flowshop instances. The predictive capacity of metrics calculated from sampled LONs is evaluated here based on the R2and RMSE indicators. Results show that the best sampling strategies provide LON metrics capable of predicting 80% and 73% of variance for small and large flowshop instances, respectively.

In the context of recommendation methods, meta-learning considers the use of previous knowledge regarding problems solution and performance to indicate the best strategy, whenever it faces a new similar problem. This paper studies the use of meta-learning to recommend local search strategies to solve several instances of permutation flowshop problems. The method proposed to conceive the meta-learning model is described considering three main phases: (i) extracting the problem features, (ii) building the performance database and (iii) training the recommendation model. In this work, we extract instances features mainly through fitness landscape analysis; build the performance data using the Irace parameter tuning algorithm and train neural networks models for recommendation. The paper also analyzes two mechanisms that support the recommendation: one using classification as its basis and another considering ranking processes. Experiments conducted on a wide range of different flowshop instances show that it is possible to recommend not only the best algorithms, but also some of their suitable configurations.

This work investigates the use of meta-learning for optimization tasks when a classic operational research problem (flowshop) is considered at the base level. It involves sequencing a set of jobs to be processed by machines in series aiming to minimize the time spent. There are various algorithms or metaheuristics proposed to solve flowshop instances and the choice of the best one usually demands time and resources. Meta-learning applied to algorithm recommendation can simulate specialists' choices as it provides a mapping between the problem characteristics (called meta-features) and the algorithm performance. This work proposes an approach for knowledge discovery operating on the performance of four metaheuristics (Hill Climbing, Tabu Search, Simulated Annealing, and Iterated Local Search) while solving several flowshop instances. Besides recommending the best metaheuristic for each instance, the proposed approach can also recommend well suited parameters values using an Irace-based training process. Despite the possibility of using complex meta-features and powerful machine learning technique, the first experiments have been conducted using simple low and high-level meta-features and a classic machine learning model called Classification and Regression Trees (CART) for the recommendation. Results show that the proposed approach is promising as the induced rules indicate that some metaheuristic parameters are preferable. Nevertheless, regarding the algorithm recommendation there is a lot of room for improvement.