Applications of Evolutionary Computation

Although numbers of heuristic algorithms are successfully developed for solving portfolio optimization problems, this is not for all cases of the large-scale ones. A large-scale portfolio optimization involves dealing with the large search space and dense variance-covariance matrix associated with the problem. This paper proposed a new multi-objective algorithm for solving a large-scale optimization problem based upon the notion of cooperative coevolutionary algorithms (CCA). The new problem decomposition scheme was designed by allowing the species-size to be dynamically adjusted as the runs progress. This scheme enhances capability of traditional CCA in dealing with non-separable optimization problem. The collaborator selection method was modified to allow the proposed CCA to perform in a multi-objective (MO) optimization framework. Additionally, the proposed algorithm, named as “DMOCCA”, was implemented for solving large-scale portfolio optimization problem with cardinality constraint using the real-world data set having scale up to 2196 dimensions. Moreover, its performances were benchmarked with those of the SPEA-II and MOPSO.

In the last few years the research community has striven to achieve a thorough understanding of the brain activity when the subject under analysis undertakes both mechanical tasks and purely mental exercises. One of the most avant-garde approaches in this regard is the discovery of connectivity patterns among different parts of the human brain unveiled by very diverse sources of information (e.g. magneto- or electro-encephalography – M/EEG, functional and structural Magnetic Resonance Imaging – fMRI and sMRI, or positron emission tomography – PET), coining the so-called brain connectomics discipline. Surprisingly, even though contributions related to the brain connectome abound in the literature, far too little attention has been paid to the exploitation of such complex spatial-temporal patterns to classify the task performed by the subject while brain signals are being registered. This manuscript covers this research niche by elaborating on the extraction of topological features from the graph modeling the brain connectivity under different tasks. By resorting to public information from the Human Connectome Project, the work will show that a selected subset of topological predictors from M/EEG connectomes suffices for accurately predicting (with average accuracy scores of up to 95%) the task performed by the subject at hand, further insights given on their predictive power when the M/EEG connectivity is inferred over different frequency bands.

Disease association studies aim at finding the genetic variations underlying complex human diseases in order to better understand the etiology of the disease and to provide better diagnoses, treatment, and even prevention. The non-linear interactions among multiple genetic factors play an important role in finding those genetic variations, but have not always been taken fully into account. This is due to the fact that searching combinations of interacting genetic factors becomes inhibitive as its complexity grows exponentially with the size of data. It is especially challenging for genome-wide association studies (GWAS) where typically more than a million single-nucleotide polymorphisms (SNPs) are under consideration. Dimensionality reduction is thus needed to allow us to investigate only a subset of genetic attributes that most likely have interaction effects. In this article, we conduct a comprehensive study by examining six widely used feature selection methods in machine learning for filtering interacting SNPs rather than the ones with strong individual main effects. Those six feature selection methods include chi-square, logistic regression, odds ratio, and three Relief-based algorithms. By applying all six feature selection methods to both a simulated and a real GWAS datasets, we report that Relief-based methods perform the best in filtering SNPs associated with a disease in terms of strong interaction effects.

For disease monitoring, grade definition and treatments orientation, specialists analyze tissue samples to identify structures of different types of cancer. However, manual analysis is a complex task due to its subjectivity. To help specialists in the identification of regions of interest, segmentation methods are used on histological images obtained by the digitization of tissue samples. Besides, features extracted from these specific regions allow for more objective diagnoses by using classification techniques. In this paper, fitness functions are analyzed for unsupervised segmentation and classification of chronic lymphocytic leukemia and follicular lymphoma images by the identification of their neoplastic cellular nuclei through the genetic algorithm. Qualitative and quantitative analyses allowed the definition of the Rényi entropy as the most adequate for this application. Images classification has reached results of 98.14% through accuracy metric by using this fitness function.

Brain Computer Interfaces provide a very challenging classification task due to small numbers of instances, large numbers of features, non-stationary problems, and low signal-to-noise ratios. Feature selection (FS) is a promising solution to help mitigate these effects. Wrapper FS methods are typically found to outperform filter FS methods, but reliance on cross-validation accuracies can be misleading due to over-fitting. This paper proposes a filter-wrapper hybrid based on Iterated Local Search and Mutual Information, and shows that it can provide more reliable solutions, where the solutions are more able to generalise to unseen data. This study further contributes comparisons over multiple datasets, something that has been uncommon in the literature.

Quantitative analysis of brain cytoarchitecture requires effective and efficient segmentation of the raw images. This task is highly demanding from an algorithmic point of view, because of the inherent variations of contrast and intensity in the different areas of the specimen, and of the very large size of the datasets to be processed. Here, we report a machine vision approach based on Convolutional Neural Networks (CNN) for the near real-time segmentation of neurons in three-dimensional images with high specificity and sensitivity. This instrument, together with high-throughput sample preparation and imaging, can lay the basis for a quantitative revolution in neuroanatomical studies.

Topoisomerases are proteins that regulate the topology of DNA by introducing transient breaks to relax supercoiling. In this paper we focus our attention on Topoisomerases 2 (TOP2), which generate double-strand DNA breaks that, if inefficiently repaired, can seriously compromise genomic stability. It is then important to gain insights on the molecular processes involved in TOP2-DNA binding. In order to do this, we collected genomic and epigenomic information from publicly available high-throughput sequencing projects and systematically quantified them within experimentally measured TOP2 binding sites. We then applied feature selection techniques in order to both increase the performance of classification and to gain insight on the particular properties that can be of biological relevance. Results obtained allowed us to identify a core set of predictive chromatin features that faithfully explain TOP2 binding.

In this paper, a new approach towards P. Polycephalum inspired computational efforts is proposed, with specific application to the problem of Multimodal transportation network design for planned cities of the future. Working with a multi-agent model of the Physarum Polycephalum, parallels are drawn between agent properties and mode characteristics, and agents are allowed to dynamically change from one mode to another. A mechanism to compare the performance of resultant multimodal networks against single mode networks involving the same component modes is demonstrated. The observations point to the potential applicability of the new approach in city planning and design.

In the context of social networks, maximizing influence means contacting the largest possible number of nodes starting from a set of seed nodes, and assuming a model for influence propagation. The real-world applications of influence maximization are of uttermost importance, and range from social studies to marketing campaigns. Building on a previous work on multi-objective evolutionary influence maximization, we propose improvements that not only speed up the optimization process considerably, but also deliver higher-quality results. State-of-the-art heuristics are run for different sizes of the seed sets, and the results are then used to initialize the population of a multi-objective evolutionary algorithm. The proposed approach is tested on three publicly available real-world networks, where we show that the evolutionary algorithm is able to improve upon the solutions found by the heuristics, while also converging faster than an evolutionary algorithm started from scratch.

Identifying Relevant Sets, i.e., variable subsets that exhibit a coordinated behavior, in complex systems is a very relevant research topic. Systems that exhibit complex dynamics are, for example, social networks, which are characterized by complex and dynamic relationships among users. A challenging topic within this context regards the identification of communities or subsets of users, both within the whole network and within specific groups. We applied the Relevance Index method, which has been shown to be effective in many situations, to the study of communities of users in the Facebook group of the Italian association of patients affected by Hidradenitis Suppurativa. Since the need for computing the Relevance Index for each possible variable subset of users makes the exhaustive computation unfeasible, we resorted to the help of an efficient niching evolutionary metaheuristic, hybridized with local searches. The communities detected through the aforementioned method have been studied to search similarities in terms of number of posts, sentiments, number of contacts, roles, behaviors, etc. The results demonstrate that it is possible to detect such subsets of users in the particular Facebook group we analyzed.

In order to better exploit scarce radio spectrum resources, the second generation of the Digital Video Broadcasting - Terrestrial standard (DVB-T2) has been developed and is under adoption in many countries, especially in Europe, for providing digital television services. The switch from the first to the second generation of DVB-T will require new operators to design their new networks and old operators to reconfigure their existing networks to better adapt to the features and opportunities of the new services. In this work, we propose an optimization model and a fast metaheuristic for the design of DVB-T2 networks. The metaheuristic is based on combining a probabilistic variable fixing procedure with an exact large neighborhood search and is developed to tackle the unsatisfying performance of state-of-the-art optimization solvers when adopted to solve realistic instances. Computational tests on realistic instances show that our metaheuristic can find solutions of much better quality than those identified by a state-of-the-art optimization solver.

A genetic algorithm for detecting a community structure in attributed graphs is proposed. The method optimizes a fitness function that combines node similarity and structural connectivity. The communities obtained by the method are composed by nodes having both similar attributes and high link density. Experiments on synthetic networks and a comparison with five state-of-the-art methods show that the genetic approach is very competitive and obtains network divisions more accurate than those obtained by the considered methods.

The dynamics of opinion formation process in a social network is of great interest for many non-equilibrium systems, such as election, competition of market share in advertising etc. By introducing local disturbance in the social network, such as the implantation of an agent, we can use numerical simulation to measure the effect of this agent on the result of the election, which has a deadline. By extending the statistical physics of damage spreading in spin models on lattice to social network, we investigate the effect of one agent on a two-party election on the time to dominance as a function of the given time to the deadline of the election. We find that certain rewiring mechanism of the social network will enhance the speed to dominance by the party that implant the agent. Using genetic algorithm, we also find good methods of rewiring that can greatly increase the efficiency of the agent. Our model is an Ising model defined on a Watts-Strogatz network. We perform Monte Carlo simulations on the effect of interaction and use a genetic algorithm with a mutation matrix to find the best way of rewiring to amplify the effect of the agent in influencing the result of the election. We also discuss the general topological feature of an optimal rewiring condition in maximizing the effect of the local disturbance in opinion formation.

In systems composed by a high number of highly coupled components, aligning the optimum of the system with the optimum of those individual components can be conflicting, especially in situations in which resources are scarce. In order to deal with this, many authors have proposed forms of biasing the optimization process. However, mostly, this works for cooperative scenarios. When resources are scarce, the components compete for them, thus those solutions are not necessarily appropriate. In this paper a new approach is proposed, in which there is a synergy between: (i) a global optimization process in which the system authority employs metaheuristics, and (ii) reinforcement learning processes that run at each component or agent. Both the agents and the system authority exchange solutions that are incorporated by the other party. The contributions are twofold: we propose a general scheme for such synergy and show its benefits in scenarios related to congestion games.

Battery energy storage systems are a key to the utilization of renewable energies, allowing for short-term storage of electricity and balancing of energy generation and consumption. However, the optimal operation of these systems is still an area of research. This paper presents operating strategies and their optimization with respect to total operational energy costs in buildings that are equipped with automated building energy management systems. The presented approach uses an evolutionary algorithm to set the parameters of the battery system controller for a rolling horizon. The combination of scheduling and control is chosen to aim at robustness against deviations of local loads from predictions. Scenarios comprising different electricity tariffs and the optimization of three operating strategies are simulated and evaluated. The results show that the operating strategies and their optimization lead to significantly different results, reflecting their ability to cope with uncertainty of future consumption and generation.

Smart grid control demands delegation of liabilities to distributed, small energy resources. Resource independent algorithm design demands abstraction from individual capabilities for integration into a general optimization model. For predictive scheduling with high penetration of renewable energy resources, agent approaches have shown good performance especially when using classifier-based decoders for modeling flexibilities. Such decoder-based methods currently are not able to cope with ensembles of individually acting energy resources. Aggregating training sets that are randomly sampled from phase-spaces of single units results in folded distributions with unfavorable properties for training a decoder. Nevertheless, for integrating e. g. a hotel, a small business, or similar with an ensemble of co-generation, heat pump, solar power, or controllable consumers a combined training set is needed. Thus, we improved the training process. We present an approach using evolution strategies for sampling ensembles that moves new instances to better positions according to spread and coverage of the feasible region. As a test case we use CMA-ES and present preliminary results demonstrating the applicability of the proposed approach.

This work aims at comparing different many-objective techniques for the optimization of mission and parallel hybrid electric power system for aircraft. In particular, this work considers, as input of the optimization, the specification of the flight mission, the size of the main components and the energy management strategy for a Medium Altitude Long Endurance Unmanned Aerial Vehicle (MALE-UAV). The goals of the optimization are maximization of electric endurance, minimization of overall fuel consumption, improvement of take-off performance and minimization of the additional volume of the hybrid electric solution with respect to the initial conventional power system. The optimization methods considered in this study are those included in the ModeFRONTIER optimization environment: NSGA-II, MOGA-II, MOSA (Multi Objective Simulated Annealing algorithm) and Evolutionary Strategy of type (µ/ρ + λ)-ES. Initially, appropriate metrics are used to compare the proposed methods in a simplified problem with only two objective functions. Then a complete optimization is performed, in order to underline the degradation of the proposed optimization methods as the size of the problem increases and to define the best method according to the number of objective functions.

We describe an algorithm to design controllers for the charging of electric vehicles. The controller is represented as a neural network, whose weights are set by an evolutionary algorithm in order to minimize the changes in the overall electrical consumption. The presented algorithm provides de-centralized controllers that also respect the privacy of the owner of electric vehicles, i.e. the controller does not share the information about charging with any third party. The presented controllers also require only a very small amount of memory and computational resources and are thus suitable for implementation in embedded systems.

The evolution of the modern power grid has evident challenges as increasing renewable distributed energy resources are outpacing grid adaptation. With increasing availability and access to real-time sensors and actuators for equipment, distributed control and optimization mechanisms are coming within technical and economic reach. Applying these now feasible mechanisms to known and existing technologies in-place brings rise to new opportunities for the integration of distributed energy resources. This work demonstrated the use of evolutionary computation in finding optimal control parameters of refrigeration systems whose dynamics are unknown and difficult to estimate. By networking evolutionary processes through administrative layers in the form of cyber-physical graph database models, controllers can be deployed at scale and then configured through genetic search algorithms and network interfaces. The premise and direction of this work focuses on leveraging relational information inferred from the graph database to improve the efficiency of the evolutionary process.

This paper describes an iterative process for generating multi-story shooter game levels by means of interlocking rooms evolved individually. The process is highly controllable by a human designer who can specify the entrances to this room as well as its size, its distribution of game objects and its architectural patterns. The small size of each room allows for computationally fast evaluations of several level qualities, but these rooms can be combined into a much larger shooter game level. Each room has two floors and is generated iteratively, with two stages of evolution and two stages of constructive post-processing. Experiments in generating an arena-based level for two teams spawning in different rooms demonstrate that the placement and allocation of entrances on each floor have a strong effect on the patterns of the final level.

Recent work has shown that genetic algorithms are a good choice for use in game design, particularly for finding improved versions of a game’s parameters to better fit a designer’s requirements. A significant issue with this approach to game optimisation is the very long time it can take to evaluate fitness, since this requires running the target game many times. In this work we test the use of several different fitness approximators, all used in a similar manner, to greatly reduce the number of times a game has to be played for the purpose of fitness evaluation. The approximators use data generated online by the genetic algorithm to train an underlying model. When the model is ready, it is invoked to provide an estimate of the fitness of each newly created individual. If this is worse than a given threshold, it is taken to be the fitness of the individual. Otherwise, the original fitness function is invoked. We assess this approach on two video games Ms PacMan and TORCS. Results are positive and move us one step closer to the goal of a games balancing tool usable in industry.

In digital games, the visual representation of game assets such as avatars or game levels can hint at their purpose, in-game use and strengths. In the Pokémon games, this is particularly prevalent with the namesake creatures’ type and the colors in their sprites. To win these games, players choose Pokémon of the right type to counter their opponents’ strengths; this makes the visual identification of type important. In this paper, computational intelligence methods are used to learn a mapping between a Pokémon’s type and its in-game sprite, colors and shape. This mapping can be useful for a designer attempting to create new Pokémon of certain types. In this paper, instead, evolutionary algorithms are used to create new Pokémon sprites by using existing color information but recombining it into a new palette. Results show that evolution can be applied to Pokémon sprites on a local or global scale, to exert different degrees of designer control and to achieve different goals.

Variants of chess have been generated in many forms and for several reasons, such as testbeds for artificial intelligence research in general game playing. This paper uses the visual properties of chess pieces as inspiration to generate new shapes for other chess-like games, targeting specific visual properties which allude to the pieces’ in-game function. The proposed method uses similarity measures in terms of pieces’ strategic role and movement in a game to identify the new pieces’ closest representatives in chess. Evolution then attempts to minimize the distance from chess pieces’ visual properties, resulting in new shapes which combine one or more chess pieces’ visual identities. While experiments in this paper focus on two chess-like games from previous publications, the method can be used for broader generation of game visuals based on functional similarities of components to known, popular games.

This work presents an evolutionary approach to optimize the parameters of a Fuzzy-based autonomous driver for the open simulated car racing game (TORCS). Using evolutionary algorithms, we intend to optimize a modular fuzzy agent designed to determine the optimal target speed as well as the steering angle during the race. The challenge in this kind of fuzzy systems is the design of the membership functions, which is usually done through a trial and error process, but in this paper an adapted real-coded Genetic Algorithm with two different fitness functions - has been applied to find the best values for these parameters, obtaining a robust design for the TORCS controller. The evolved drivers were tested and evaluated competing against other TORCS controllers in practice mode, without rivals, and real races. The optimized fuzzy-controllers yield a very good performance, mainly in tracks that have many turning points, which are, in turn, the most difficult for any autonomous agent. Thus, this is a real enhancement of the baseline fuzzy controllers which had several difficulties to drive in this kind of circuits.

Monte-Carlo Tree Search (MCTS) has shown particular success in General Game Playing (GGP) and General Video Game Playing (GVGP) and many enhancements and variants have been developed. Recently, an on-line adaptive parameter tuning mechanism for MCTS agents has been proposed that almost achieves the same performance as off-line tuning in GGP.In this paper we apply the same approach to GVGP and use the popular General Video Game AI (GVGAI) framework, in which the time allowed to make a decision is only 40 ms. We design three Self-Adaptive MCTS (SA-MCTS) agents that optimize on-line the parameters of a standard non-Self-Adaptive MCTS agent of GVGAI. The three agents select the parameter values using Naïve Monte-Carlo, an Evolutionary Algorithm and an N-Tuple Bandit Evolutionary Algorithm respectively, and are tested on 20 single-player games of GVGAI.The SA-MCTS agents achieve more robust results on the tested games. With the same time setting, they perform similarly to the baseline standard MCTS agent in the games for which the baseline agent performs well, and significantly improve the win rate in the games for which the baseline agent performs poorly. As validation, we also test the performance of non-Self-Adaptive MCTS instances that use the most sampled parameter settings during the on-line tuning of each of the three SA-MCTS agents for each game. Results show that these parameter settings improve the win rate on the games Wait for Breakfast and Escape by 4 times and 150 times, respectively.

Deceptive games are games where the reward structure or other aspects of the game are designed to lead the agent away from a globally optimal policy. While many games are already deceptive to some extent, we designed a series of games in the Video Game Description Language (VGDL) implementing specific types of deception, classified by the cognitive biases they exploit. VGDL games can be run in the General Video Game Artificial Intelligence (GVGAI) Framework, making it possible to test a variety of existing AI agents that have been submitted to the GVGAI Competition on these deceptive games. Our results show that all tested agents are vulnerable to several kinds of deception, but that different agents have different weaknesses. This suggests that we can use deception to understand the capabilities of a game-playing algorithm, and game-playing algorithms to characterize the deception displayed by a game.

Convolutional highways are based on multiple stacked convolutional layers for feature preprocessing. Like many other convolutional networks convolutional highways are parameterized by numerous hyperparameters that have to be tuned carefully. We introduce an evolutionary algorithm (EA) for optimization of the structure and tuning of hyperparameters of convolutional highways and demonstrate the potential of this optimization setting on the well-known MNIST data set. The EA employs Rechenberg’s mutation rate control and a niching mechanism to overcome local optima. An experimental study shows that the EA is capable of evolving convolutional highway networks from scratch with only few evaluations but achieving competitive accuracy. Further, the EA is able to significantly improve standard network configurations.

Learning Classifier Systems (LCSs) have demonstrated their classification capability by employing a population of polymorphic rules in addressing numerous benchmark problems. However, although the produced solution is often accurate, the alternative ways to represent the data in a single population obscure the underlying patterns of a problem. Moreover, once a population is dominated by over-general rules, the system will sink into the local optimal trap. To grant a problem’s patterns more transparency, the redundant rules and optimal rules need to be distinguished. Therefore, the bagging method is introduced to LCSs with the aim to reduce the variance associated with redundant rules. A novel rule reduction method is proposed to reduce the rules’ polymorphism in a problem. This is tested with complex binary problems with typical epistatic, over-lapping niches, niche-imbalance, and specific-addiction properties at various scales. The results show the successful highlighting of the patterns for all the tested problems, which have been addressed successfully. Moreover, by combining the boosting method with LCSs, the hybrid system could adjust previously defective solutions such that they now represent the correct classification of data.

Feature extraction is an essential process for image data dimensionality reduction and classification. However, feature extraction is very difficult and often requires human intervention. Genetic Programming (GP) can achieve automatic feature extraction and image classification but the majority of existing methods extract low-level features from raw images without any image-related operations. Furthermore, the work on the combination of image-related operators/descriptors in GP for feature extraction and image classification is limited. This paper proposes a multi-layer GP approach (MLGP) to performing automatic high-level feature extraction and classification. A new program structure, a new function set including a number of image operators/descriptors and two region detectors, and a new terminal set are designed in this approach. The performance of the proposed method is examined on six different data sets of varying difficulty and compared with five GP based methods and 42 traditional image classification methods. Experimental results show that the proposed method achieves better or comparable performance than these baseline methods. Further analysis on the example programs evolved by the proposed MLGP method reveals the good interpretability of MLGP and gives insight into how this method can effectively extract high-level features for image classification.

In classification and clustering problems, selecting a subset of discriminative features is a challenging problem, especially when hundreds or thousands of features are involved. In this framework, Evolutionary Computation (EC) techniques have received a growing scientific interest in the last years, because they are able to explore large search spaces without requiring any a priori knowledge or assumption on the considered domain. Following this line of thought, we developed a novel strategy to improve the performance of EC-based algorithms for feature selection. The proposed strategy requires to rank the whole set of available features according to a univariate evaluation function; then the search space represented by the first M ranked features is searched using an evolutionary algorithm for finding feature subsets with high discriminative power. Results of comparisons demonstrated the effectiveness of the proposed approach in improving the performance obtainable with three effective and widely used EC-based algorithm for feature selection in high dimensional data problems, namely Ant Colony Optimization (ACO), Particle Swarm Optimization (PSO) and Artificial Bees Colony (ABC).

This paper presents and describes CGP4Matlab, a powerful toolbox that allows to run Cartesian Genetic Programming within MATLAB. This toolbox is particularly suited for signal processing and image processing problems. The implementation of CGP4Matlab, which can be freely downloaded, is described. Some encouraging results on the problem of pitch estimation of musical piano notes achieved using this toolbox are also presented. Pitch estimation of audio signals is a very hard problem with still no generic and robust solution found. Due to the highly flexibility of CGP4Matlab, we managed to apply a new cartesian genetic programming based approach to the problem of pitch estimation. The obtained results are comparable with the state of the art algorithms.

The Relevance Index (RI) is an information theory-based measure that was originally defined to detect groups of functionally similar neurons, based on their dynamic behavior. More in general, considering the dynamical analysis of a generic complex system, the larger the RI value associated with a subset of variables, the more those variables are strongly correlated with one another and independent from the other variables describing the system status. We describe some early experiments to evaluate whether such an index can be used to extract relevant feature subsets in binary pattern classification problems. In particular, we used a PSO variant to efficiently explore the RI search space, whose size equals the number of possible variable subsets (in this case $$2^{104}$$) and find the most relevant and discriminating feature subsets with respect to pattern representation. We then turned such relevant subsets into a new smaller set of richer features, whose values depend on the values of the binary features they include. The paper reports some exploratory results we obtained in a simple character recognition task, comparing the performance of RI-based feature extraction and selection with other classical feature selection/extraction approaches.

Super-resolution reconstruction (SRR) allows for producing a high-resolution (HR) image from a set of low-resolution (LR) observations. The majority of existing methods require tuning a number of hyper-parameters which control the reconstruction process and configure the imaging model that is supposed to reflect the relation between high and low resolution. In this paper, we demonstrate that the reconstruction process is very sensitive to the actual relation between LR and HR images, and we argue that this is a substantial obstacle in deploying SRR in practice. We propose to search the hyper-parameter space using a genetic algorithm (GA), thus adapting to the actual relation between LR and HR, which has not been reported in the literature so far. The results of our extensive experimental study clearly indicate that our GA improves the capacities of SRR. Importantly, the GA converges to different values of the hyper-parameters depending on the applied degradation procedure, which is confirmed using statistical tests.

Feature extraction is the first step in building real-life classification engines—it aims at elaborating features to characterize objects that are to be labeled by a trained model. Time-consuming feature extraction requires domain expertise to effectively design features. Deep neural networks (DNNs) appeared as a remedy in this context—their shallow layers perform representation learning, being an automated discovery of various-level features that robustly represent objects. However, the representations that are being learnt are still extremely difficult to interpret, and DNNs are prone to memorizing small datasets. In this paper, we introduce evolvable deep features (EDFs)—a DNN is used to extract automatic features that undergo genetic feature selection. Such evolved features are fed into a supervised learner. The experiments, backed up with statistical tests, performed on multi- and binary-class sets showed that our approach automatically learns object representations, greatly reduces the number of features without deteriorating the performance of trained models, and can even boost their classification performance.

Object recognition is a widely studied problem in computer vision. Template matching with correlation filters is one of the most accurate strategies for target recognition. However, it is computationally expensive, particularly when there is no restriction in the pose of the object of interest and an exhaustive search is implemented. This work proposes the use of a Covariance Matrix Adaptation Evolution Strategy (CMA-ES) for post-processing template matched filters. The proposed strategy searches for the best template matching guided by the discrimination capability of a correlation-based filter, considering a vast set of filters. CMA-ES is used to find the best match and determine the correct pose or orientation parameters of a target object. The proposed method demonstrates that CMA-ES is effective for multidimensional problems in a huge search space, which makes it a suitable candidate for target recognition in unconstrained applications. Experimental results show high efficiency in terms of the number of function evaluations and locating the correct pose parameters based on the DC measure.

EvoFilter is an evolutionary algorithm based tool for searching through large non-dominated fronts in order to find a subset of solutions that are of interest to the user. EvoFilter is designed to take the output of existing Multi Objective Evolutionary Algorithms and act as a decision support tool for users. Currently EvoFilter is available for all to use on-line [1]. This paper evaluates the performance of EvoFilter by creating a large number of randomised filter specifications which are then applied using EvoFilter and a simple filter to a range of non-dominated fronts created by a portfolio of Multi Objective Genetic Algorithms (MOGAs). The results show that EvoFilter is capable of finding sets of solutions that meet the users’ requirements more closely than those found using the simple filter. EvoFilter increases performance on some objectives by including relevant solutions event if these solutions slightly lessen performance on other objectives. The filter discussed in this paper may be accessed at [1].

Find Your Doctor (FYD) aims at becoming the first Job-placement agency in Italy dedicated to PhDs who are undergoing the transition outside Academia. To support the FYD Human Resources team we started a research project aimed at extracting, from texts (questionnaires) provided by a person telling his/her experience, a set of well defined soft skills. The final aim of the project is to produce a list of researchers ranked w.r.t. their degree of soft skills ownership. In the context of this project, this paper presents an approach employing machine learning techniques aimed at classifying the researchers questionnaires w.r.t. a pre-defined soft skills taxonomy. This paper also presents some preliminary results obtained in the “communication” area of the taxonomy, which are promising and worth of further research in this direction.

A visualisation method is presented that is intended to assist evolutionary algorithm users with the parametrisation of their algorithms. The visualisation method presents the convergence and diversity properties such that different parametrisations can be easily compared, and poor performing parameter sets can be easily identified and discarded. The efficacy of the visualisation is presented using a set of benchmark optimisation problems from the literature, as well as a benchmark water distribution network design problem. Results show that it is possible to observe the different performance caused by different parametrisations. Future work discusses the potential of this visualisation within an online tool that will enable a user to discard poor parametrisations as they execute to free up resources for better ones.

This contribution presents a metaheuristic approach that integrates evolution strategies into genetic algorithms using a fuzzy rule based inference system to evaluate schedules in a generalized steelmaking and continuous casting production system. The genetic algorithm controls the job sequences assigned to the machines while the setting of jobs initial processing dates at the converter are optimize by means of evolution strategies. The fuzzy inference system gives an overall evaluation of the schedule quality by controlling discontinuities and transit times with different degrees of acceptance throughout the evolution process. This approach integrates an embedded search procedure to overcome one of the weaknesses of metaheuristic scheduling methods of setting initial dates for task processing and is especially suited for highly nonlinear objective functions as in this case. A general structure of the steelmaking and continuous casting production system is consider with an arbitrary number of machines at each stage, with production of several steel grades and types (e.g. slabs and billets). Technological constraints such as continuous casting between jobs (batches) and in process time of liquid steel are included. For illustration purposes, a real sized problem is solve.

In many industrial problem domains, when faced with a combinatorial optimisation problem, a “good enough, quick enough” solution to a problem is often required. Simple heuristics often suffice in this case. However, for many domains, a simple heuristic may not be available, and designing one can require considerable expertise. Noting that a wide variety of problems can be represented as graphs, we describe a system for the automatic generation of constructive heuristics in the form of Python programs by mean of grammatical evolution. The system can be applied seamlessly to different graph-based problem domains, only requiring modification of the fitness function. We demonstrate its effectiveness by generating heuristics for the Travelling Salesman and Multi-Dimensional Knapsack problems. The system is shown to be better or comparable to human-designed heuristics in each domain. The generated heuristics can be used ‘out-of-the-box’ to provide a solution, or to augment existing hyper-heuristic algorithms with new low-level heuristics.

This paper presents an experimental study on the efficacy of a rotation-invariant Differential Evolution (based on current-to-rand mutation) on a benchmark of test problems in its non-rotated and rotated version. Numerical results show that standard Differential Evolution outperforms rotation-invariant Differential Evolution on the benchmark under consideration for both non-rotated and rotated problems. In other words, the rotation-invariant Differential Evolution does not seem to be more efficient than its standard counterpart to address rotated problems. According to our interpretation, these experimental results show that rotated problems are simply different problems with respect to the non-rotated problems. Furthermore, rotation-invariant Differential Evolution is characterised by its moving operator: it generates an offspring by perturbing all the design variables of a candidate solution at the same time. This logic does not appear to guarantee a better performance on rotated problems.

We propose the Multi-strategy Differential Evolution (MsDE) algorithm to construct and maintain a self-adaptive ensemble of search strategies while solving an optimization problem. The ensemble of strategies is represented as agents that interact with the candidate solutions to improve their fitness. In the proposed algorithm, the performance of each agent is measured so that successful strategies are promoted within the ensemble. We propose two performance measures, and show their effectiveness in selecting successful strategies. We then present three population adaptation mechanisms, based on sampling, clone-best and clone-multiple adaptation schemes. The MsDE with different performance measures and population adaptation schemes is tested on the CEC2013 benchmark functions and compared with basic DE and with Self-Adaptive DE (SaDE). Our results show that MsDE is capable of efficiently adapting the strategies and parameters of DE and providing competitive results with respect to the state-of-the-art.

A generic framework for incorporating constraint handling techniques (CHTs) into multi-objective evolutionary algorithms (MOEAs) is proposed to resolve the differences between MOEAs from algorithmic and implementation perspective with respect to the incorporation of CHTs. To verify the effectiveness of the proposed framework, the performances of the combined algorithms of five CHTs and four MOEAs on eight constrained multi-objective optimization problems are investigated with the proposed framework. The experimental results show that the outperforming CHT can vary by constrained multi-objective optimization problems, as far as examined in this study.

This paper exposes a new hybrid approach based on Ant Colony Optimization heuristics, Route First-Cluster Second methods and Local search procedures, combined to generate high quality solutions for the Vehicle Routing Problem. This method uses the parallel computing power of modern general purpose GPUs and multicore CPUs, outperforming current ACO-based VRP solvers and showing to be a competitive approach compared to other high performing metaheuristic solvers.

Neuroevolution represents a growing research field in Artificial and Computational Intelligence. The adjustment of the network weights and the topology is usually based on a single performance criterion. Approaches that allow to consider several – potentially conflicting – criteria are only rarely taken into account.This paper develops a novel combination of the NeuroEvolution of Augmenting Topologies (NEAT) algorithm with modern indicator-based evolutionary multi-objective algorithms, which enables the evolution of artificial neural networks for multi-objective tasks including a large number of objectives. Several combinations of evolutionary multi-objective algorithms and NEAT are introduced and discussed. The focus lies on variants with modern indicator-based selection since these are considered as efficient methods for higher dimensional tasks. This paper presents the first combination of these algorithms and NEAT. The experimental analysis shows that the novel algorithms are very promising for multi-objective Neuroevolution.

Developing robotic systems that can evolve in real-time and real-space is a long term objective with technological as well as algorithmic milestones on the road. Technological prerequisites include advanced 3D-printing, automated assembly, and robust sensors and actuators. The necessary evolutionary mechanisms need not wait for these, they can be developed and investigated in simulations. In this paper, we present a system to simulate online evolution of constructible robots, where (1) the population members (robots) concurrently exist and evolve their morphologies and controllers, (2) all robots can be physically constructed. Experiments with this simulator provide us with insights into differences of using online and offline evolutionary setups.

We present a study on morphological traits of evolved modular robots. We note that the evolutionary search space –the set of obtainable morphologies– depends on the given representation and reproduction operators and we propose a framework to assess morphological traits in this search space regardless of a specific environment and/or task. To this end, we present eight quantifiable morphological descriptors and a generic novelty search algorithm to produce a diverse set of morphologies for any given representation. With this machinery, we perform a comparison between a direct encoding and a generative encoding. The results demonstrate that our framework permits to find a very diverse set of bodies, allowing a morphological diversity investigation. Furthermore, the analysis showed that despite the high levels of diversity, a bias to certain traits in the population was detected. Surprisingly, the two encoding methods showed no significant difference in the diversity levels of the evolved morphologies or their morphological traits.

Four-legged mammals are capable of showing a great variety of movement patterns, ranging from a simple walk to more complex movement such as trots and gallops. Imbuing this diversity to quadruped robots is of interest in order to improve both mobility and reach. Within the field of Evolutionary Robotics, Quality Diversity techniques have shown a remarkable ability to produce not only effective, but also highly diverse solutions. When applying this approach to four-legged robots an initial problem is to create viable movement patterns that do not fall. This difficulty stems from the challenging fitness gradient due to the mammalian morphology. In this paper we propose a solution to overcome this problem by implementing incremental evolution within the Quality Diversity framework. This allows us to evolve controllers that become more complex while at the same time utilizing the diversity produced by Quality Diversity. We show that our approach is able to generate high fitness solutions early in the search process, keep these solutions and perform a more open-ended search towards the end of evolution.

Automated design of swarm behaviors with a top-down approach is a challenging research question that has not yet been fully addressed in the robotic swarm literature. This paper seeks to explore the possibility of using an evolutionary algorithm to evolve, rather than hand code, a wide repertoire of behavior primitives enabling more effective control of a large group or swarm of unmanned systems. We use the MAP-elites algorithm to generate a repertoire of controllers with varying abilities and behaviors allowing the swarm to adapt to user-defined preferences by selection of a new appropriate controller. To test the proposed method we examine two example applications: perimeter surveillance and network creation. Perimeter surveillance require agents to explore, while network creation requires them to disperse without losing connectivity. These are distinct application that have drastically different requirements on agent behavior, and are a good benchmark for our swarm controller optimization framework. We show a performance comparison between a simple weighted controller and a parametric controller. Evolving controllers allows for specifying desired behaviors top-down, in terms of objectives to solve, rather than bottom-up.

Recent developments in memory-augmented neural networks allowed sequential problems requiring long-term memory to be solved, which were intractable for traditional neural networks. However, current approaches still struggle to scale to large memory sizes and sequence lengths. In this paper we show how access to an external memory component can be encoded geometrically through a novel HyperNEAT-based Neural Turing Machine (HyperNTM). The indirect HyperNEAT encoding allows for training on small memory vectors in a bit vector copy task and then applying the knowledge gained from such training to speed up training on larger size memory vectors. Additionally, we demonstrate that in some instances, networks trained to copy nine bit vectors can be scaled to sizes of 1,000 without further training. While the task in this paper is simple, the HyperNTM approach could now allow memory-augmented neural networks to scale to problems requiring large memory vectors and sequence lengths.

Client-side Javascript execution environments (browsers) allow anonymous functions and event-based programming concepts such as callbacks. We investigate whether a mutate-and-test approach can be used to optimise web page load time in these environments. First, we characterise a web page load issue in a benchmark web page and derive performance metrics from page load event traces. We parse Javascript source code to an AST and make changes to method calls which appear in a web page load event trace. We present an operator based solely on code deletion and evaluate an existing “community-contributed” performance optimising code transform. By exploring Javascript code changes and exploiting combinations of non-destructive changes, we can optimise page load time by 41% in our benchmark web page.

The Late Acceptance Hill-Climbing (LAHC) algorithm is a one-point search meta-heuristic with a single parameter. Like Simulated Annealing (SA) it sometimes accepts worsening moves, however it is far more simple and does not require complex parameter setting. In this paper we study an application of LAHC to the Combinatorial Interaction Testing (CIT) problem. CIT is a cost-effective black-box sampling technique for discovering interaction faults in highly configurable systems. There are several techniques for CIT; one of the most established and well-known is Covering Arrays by Simulated Annealing (CASA). CASA is a layered search framework using SA in its most inner layer. Here we replace SA in CASA with LAHC, proposing a modified framework, Covering Arrays by Late Acceptance (CALA). Our experimental evaluation demonstrates that LAHC yields better or equal quality solutions compared to SA for all but one of the 35 benchmark instances tested.

Multicore processors have now become the norm. However, for many embedded real-time systems their use introduces challenges in verification as their shared components are potential channels for interference. Of particular interest is the determination for each task of its worst case (longest) execution time (WCET). In this paper, we investigate the effectiveness of a variety of metaheuristic search algorithms for dynamically finding extreme execution times of tasks executing on a multicore processor. Over finite search spaces, these are shown to perform considerably better than randomly generated test inputs and the work reveals significant performance differences between the various algorithms.

A way to deal with uncertainties in the fitness function of an optimization problem is robust optimization, which optimizes the expected value of the fitness. In the context of evolutionary optimization, it is a common practice to compute the expected value of the fitness approximately with the help of Monte-Carlo simulation. This approach requires a lot of evaluations of the fitness function in order to evaluate an individual and thus it can be very compute-intensive.In the present paper, we propose a coevolution-based approach for the robust optimization of problems with a fitness function basically depending on discrete random variables, which conditionally depend on the decision variables. Experiments on three benchmark functions show that the approach yields a good trade-off between the number of required fitness function evaluations and the quality of the results.

Dynamic constrained optimization problems have received increasing attention in recent years. We study differential evolution which is one of the high performing class of algorithms for constrained continuous optimization in the context of dynamic constrained optimization. The focus of our investigations are repair methods which are crucial when dealing with dynamic constrained problems. Examining recently introduced benchmarks for dynamic constrained continuous optimization, we analyze different repair methods with respect to the obtained offline error and the success rate in dependence of the severity of the dynamic change. Our analysis points out the benefits and drawbacks of the different repair methods and gives guidance to its applicability in dependence on the dynamic changes of the objective function and constraints.

Evolutionary algorithms (EAs) are a good choice to solve dynamic optimization problems. Objective functions changing over time are challenging because after a change the EA has to adapt its population to find the new optimum. Prediction techniques that estimate the position of the next optimum can be incorporated into the EA. After a change, the predicted optimum can be employed to move the EA’s population to a promising region of the solution space in order to accelerate convergence and improve accuracy in tracking the optimum. In this paper we introduce a recurrent neural network-based prediction approach. In an experimental study on the Moving Peaks Benchmark and dynamic variants of the Sphere, Rosenbrock, and Rastrigin functions we compare it to an autoregressive prediction approach and an EA without prediction. The results show the competitiveness of our approach and its suitability especially for repeated optima.

Dynamic optimization problems (DOPs) are problems that change over time and many real-world problems are classified as DOPs. However, most of investigations in this domain are focused on tracking moving optima (TMO) without considering any other objectives which creates a gap between real-world problems and academic research in this area. One of the important optimization objectives in many real-world problems is previous-solution displacement restriction (PSDR) in which successive solutions should not be much different. PSDRs can be categorized as a multi-objective problem in which the first objective is optimality and the second one is minimizing the displacement of consecutive solutions which also can represents switching cost. Moreover, PSDRs are counted as dynamic time-linkage problems (DTPs) because the current chosen solution by the optimizer will change the next search space. In this paper, we propose a new hybrid method based on particle swarm optimization (PSO) for PSDRs based on their characteristics. The experiments are done on moving peaks benchmark (MPB) and the performance of the proposed algorithm alongside two comparison ones are investigated on it.

Characterization of dynamism is an important issue for utilizing or tailoring of several dynamic multi-objective evolutionary algorithms (DMOEAs). One such characterization is the change detection, which is based on proposing explicit schemes to detect the points in time when a change occurs. Additionally, detecting severity of change and incorporating with the DMOEAs is another attempt of characterization, where there is only a few related works presented in the literature. In this paper, we propose a type-detection mechanism for dynamic multi-objective optimization problems, which is one of the first attempts that investigate the significance of type detection on the performance of DMOEAs. Additionally, a hybrid technique is proposed which incorporates our type detection mechanism with a given DOMEA. We present an empirical evaluation by using seven test problems from all four types and five performance metrics, which clearly validate the motivation of type detection as well as significance of our hybrid technique.

This paper aims to present and evaluate a software that uses an evolutionary strategy to design weekly nutritional menus for School Feeding. The software ensures the nutritional needs of students and also minimizes the total cost of the menu. We based our nutritional needs on the Brazilian National School Feeding Programme (PNAE). This program takes into account: (i) the age of the student; (ii) some preparations issues as color, consistency and, variety; and also (iii) the maximum amount to be paid per meal. Our software generates, in less than five minutes, a set of menus, and the nutritionist can choose the menu that suits his/her best. We evaluate our algorithm using the Weighted-Sum approach, and our results show that the obtained 5-days menus using the proposed methodology not only comply with the restrictions imposed by the authorities but also produce inexpensive and healthy menus. We also appraise the software itself using an opinion pool among nine nutritionists. The professionals considered our software above expectations.