A current challenge in computational technology is the increasing energy consumption associated with the development and deployment of domain-specific large language models (LLMs), such as those used in the medical domain. In this work, we explore the balance between performance and energy consumption trade-offs across several small-scale LLMs (1B–8B parameters), fine-tuned for medical multiple choice question answering using efficient methods such as Low-Rank Adaptation (LoRA) and quantization. For our experiments, we utilized two entry-level professional hardware setups representative of accessible workstation environments, measuring energy and time consumption across multiple models. The results show that moderate-sized models, such as LLaMA-3.1-8B, can achieve accuracy levels comparable to much larger biomedical-tuned models, such as PMC-LLaMA-13B, while requiring substantially less hardware. Additionally, smaller models, such as the 1B and 3B parameter versions, demonstrate notable efficiency during inference, making them suitable for deployment on edge devices, which are traditionally highly constrained by energy and computational resources. This paper offers practical insights into the deployment of medical models under realistic hardware limitations, supporting the goals of energy-aware and accessible machine learning.

Evolution is a powerful problem-solving technique, extensively used for designing racing car controllers, but with a series of challenges: an evaluation function that can separate the best controllers from the rest, and a series of operators that can explore different possibilities in the controller search space. Within the context of the TORCS racing simulator, in this article, we introduce a selection policy based on competition calledGrand Prix Selection(GPS), which will be able to increase robustness by using something more realistic than solo race scores to select individuals. Additionally, we increase the exploitative power of this kind of selection via a BLX operator with continuously decreasing$alpha$. We compare these new selection and operator with hybrid approaches that apply GPS only part of the time, as well as other classical crossover operators. In general, experiments show that these combined improvements establish a new level of performance of evolved controllers, being able to beat both standard and previously evolved ones, as well as a high-ranked controller of TORCS competitions.

Synchronous operation is not the most natural, as in biologically inspired, mode to run distributed algorithms. In many grid, cloud or volunteer setups nodes are heterogeneous, or simply are not available at the exact same time; this is a challenge for the researcher if their full performance is going to be actually leveraged. Asynchronous distributed evolutionary algorithms try to solve this by dropping the homogeneity, as well as the synchronicity, assumption. These algorithms share the population between distributed workers which execute the actual evolutionary process by taking samples of the population, and replacing them in the population pool by evolved individuals. The performance of these EAs depends in part on the selection of parameters for the EA running in each worker. In this paper we study how randomly varying parameters in distributed evolutionary algorithms affects performance. Experiments were conducted in the AWS cloud using 2, 6 and 12 virtual machine configurations, with both homogeneous and heterogeneous random settings using five test functions for real-valued optimization and the OneMax binary problem. The results suggest that this method can produce a performance that is competitive with instances of the algorithm using workers with parameters specially tuned for the benchmark.

The term moving target defense or MTD describes a series of techniques that change the configuration of an Internet-facing system; in general, the technique consists of changing the visible configuration to avoid offering a fixed target to service profiling techniques. Additionally, configurations need to be as secure as possible and, since change needs to be frequent, to generate also as many as possible. We previously introduced a proof of concept where we used a simplified evolutionary algorithm for generating these configurations. In this paper we improve this algorithm, trying to adapt it to the specific characteristics of the fitness landscape, and also looking at finding as many solutions as possible.

Abstract Designing a story is widely considered a crafty yet critical task that requires deep specific human knowledge in order to reach a minimum quality and originality. This includes designing at a high level different elements of the film; these high‐level elements are called tropes when they become patterns. The present paper proposes and evaluates a methodology to automatically synthesize sets of tropes in a way that they maximize the potential rating of a film that conforms to them. We use machine learning to create a surrogate model that maps film ratings from tropes, trained with the data extracted and processed from huge film databases in Internet, and then we use a genetic algorithm that uses that surrogate model as evaluator to optimize the combination of tropes in a film. In order to evaluate the methodology, we analyse the nature of the tropes and their distributions in existing films, the performance of the models and the quality of the sets of tropes synthesized. The results of this proof of concept show that the methodology works and is able to build sets of tropes that maximize the rating and that these sets are genuine. The work has revealed that the methodology and tools developed are directly suitable for assisting in the plots generation as an authoring tool and, ultimately, for supporting the automatic generation of stories, for example, in massively populated videogames.

One of the main problems in the design through optimization of car racing bots is the inherent noise in the optimization process: besides the fact that the fitness is a heuristic based on what we think are the keys to success and as such just a surrogate for the ultimate objective, winning races, fitness itself is uncertain due to the stochastic behavior of racing conditions and the rest of the (simulated) racers. The fuzzy-based genetic controller for the car racing simulator TORCS we have defined in previous works is based on two fuzzy sub-controllers, one for deciding on the wheel steering angle and another to set the car target speed at the next simulation tick. They are both optimized by means of an Evolutionary Algorithm, which considers an already tested fitness function focused on the maximization of the average speed during the race and the minimization of the car damage. The noisy environment asks for keeping diversity high during evolution, that is why we have added a Blend Crossover (BLX-α) operator, which is, besides, able to exploit current results at the same time it explores. Additionally, we try to address uncertainty in selection by introducing a novel selection policy of parents based in races, where the individuals are grouped and compete against others in several races, so just the firsts ranked will remain in the population as parents. Several experiments have been conducted, testing the value of the different controllers. The results show that the combination of a dynamic BLX-α crossover operator plus the pole position selection policy clearly beats the rest of approaches. Moreover, in the comparison of this controller with one of the participants of the prestigious international Simulated Car Racing Championship, our autonomous driver obtains much better results than the opponent.

In this work, we present an original method based on evolutionary algorithms for generating basic structures for the physics-based game Angry Birds, with the ultimate objective of creating Angry Birds levels with the minimum number of constraints. We set out to evolve free-form structures, and this means searching in a larger space. In this paper, we test how using a physics engine enables us to evaluate much more levels than a game engine simulation. In order to do this, we compare the results of experiments using both types of simulators and propose fitness functions accordingly. Results show the execution time drastically drops from 5 hours to less than 20 minutes on average.

This paper investigates the emergent properties of a self-organized dynamic network for structured population-based metaheuristics. The system displays complex and emergent behavior whose most visible trait is the self-organization of the population into dynamic clusters. Furthermore, relevant variables that describe the system display 1/f noise, which is a characteristic of many complex systems. These properties were previously detected with a time-invariant population (i.e., individuals with fixed fitness vales). In this work, the investigation is extended to dynamic populations (time-varying fitness values), a scenario that models more accurately the behavior of population-based metaheuristics. Several types of fitness variation rules were tested. The experiments show that dynamic populations also display the self-organizing properties and the behavioral patterns of the stationary fitness version, as long as the intensity of the changes is kept below a certain level.

International audience

Interactive evolution, where users' preferences guide the search, is one of the techniques employed by Evolutionary Art researchers. It can be implemented as a web application to lower the access threshold since it often depends on volunteers who visit the system for fitness assignment. However, several drawbacks limit user participation: human fatigue and boredom result from evaluating a large number of phenotypes. To tackle these issues, in this paper we propose an IEC system designed using a human-centered approach, with a framework consisting of a social network of volunteers interacting with a population also consisting of a network of phenotypes. The use of a graph model is proposed as a practical and efficient tool for mapping the relationships between actors and objects in the system. A case study is presented as proof-of-concept, providing both conceptual and implementation details of the graph model as it is applied in the implementation of an IEC system. Our experiments show that the data model can be successfully used to implement a gamification technique developed to increase users' engagement, which implies that this technique can be successfully used to decrease user fatigue and thus increase the performance of the interactive system.

The software presented in this paper is an implementation of an intuitionistic fuzzy inference system. Such type of fuzzy inference systems provide an extra layer of uncertainty, called indeterminacy, that the user can integrate in the antecedents and consequents of the fuzzy system. The additional calculations required to make an inference in this type of system needs a negligible extra amount of computational resources, making it a low-cost alternative to type-2 fuzzy inference systems. At the current time, no other implementation of such type of system exist that is open source and free of charge. The software is developed in Clojure in order to leverage the Java libraries, the JVM itself, and the capabilities of the programming language to implement concurrency in a convenient manner. However, the goal of this implementation is to provide a language-agnostic interface based in a REST API, which can be used by any programming language capable of handling HTTP requests. A comparison between a traditional type-1 fuzzy inference is provided, where the reader can observe how the indeterminacy affects the outputs of the system.

It is not usual practice in the evolutionary algorithms area to benchmark different operations in order to choose the best language for a single or multilanguage implementation. Researchers rely instead on common practice or frameworks using mainstream languages. That is why it is usual practice to choose compiled languages (namely Java or C/C++) when implementing evolutionary algorithms, without considering other languages or rejecting them outright on the basis of performance. Since there is a myriad of languages nowadays, we considered it an interesting challenge to measure their speed when performing frequent operations in evolutionary algorithms. In this paper we have tested three basic evolutionary algorithm operations over binary chromosomes: bitflip mutation, crossover and the OneMax fitness function. As a performance measure, the speed for both popular and not so popular computer languages have been used. In general, the results confirm that compiled languages scale and perform better, but also in some cases have a behaviour that is independent of the size of the chromosome. Additionally, results show that other languages, such as Go (compiled) or Python (interpreted) are fast enough for most purposes. Besides, these experiments show which of these operations are, in fact, the best for choosing an implementation language based on its performance.

Using volunteer's browsers as a computing resource presents several advantages, but it remains a challenge to fully harness the browser's capabilities and to model the user's behavior so that those capabilities can be leveraged optimally. These are the objectives of this paper, where we present the results of several evolutionary computation experiments with different implementations of a volunteer computing framework called NodIO, designed to be easily deployable on freely available cloud resources. We use different implementations to find out which one is able to get the user to lend more computing cycles and test different problems to check the influence it has on said performance, as measured by the time needed to find a solution, but also by the number of users engaged. From these experiments we can already draw some conclusions, besides the fact that volunteer computing can be a valuable computing resource and that it is essential to be as open as possible with software and data: the user has to be kept engaged to obtain as many computing cycles as possible, the client has to be built to use the computer capabilities fully, and, finally, that the user contributions follow a common statistical distribution.

In the last year, thanks to the Ms. Pac-Man Versus Ghosts Competition, the game of Ms. Pac-Man has gained increasing attention from academics in the field of computational intelligence. In this paper, we contribute to this research stream by presenting a simple genetic algorithm with lexicographic ranking (GALR) for the optimization of flocking strategy-based ghost controllers. Flocking strategies are a paradigm for intelligent agents characterized by showing emergent behavior and for having very little computational and memory requirements, making them well suited for commercial applications and mobile devices. In particular, we study empirically the effect of optimizing homogeneous and heterogeneous teams. The computational analysis shows that the flocking strategy-based controllers generated by the proposed GALR outperform the ghost controllers included in the competition framework and some of those presented in the literature.

MOTIVATION: Self-organizing maps (SOMs) are readily available bioinformatics methods for clustering and visualizing high-dimensional data, provided that such biological information is previously transformed to fixed-size, metric-based vectors. To increase the usefulness of SOM-based approaches for the analysis of genomic sequence data, novel representation methods are required that automatically and objectively transform aligned nucleotide sequences into numeric vectors, dealing with both nucleotide ambiguity and gaps derived from sequence alignment. RESULTS: Six different codification variants based on Euclidean space, just like SOM processing, have been tested using two SOM models: the classical Kohonen's SOM and growing cell structures. They have been applied to two different sets of sequences: 32 sequences of small sub-unit ribosomal RNA from organisms belonging to the three domains of life, and 44 sequences of the reverse transcriptase region of the pol gene of human immunodeficiency virus type 1 belonging to different groups and sub-types. Our results show that the most important factor affecting the accuracy of sequence clustering is the assignment of an extra weight to the presence of alignment-derived gaps. Although each of the codification variants shows a different level of taxonomic consistency, the results are in agreement with sequence-based phylogenetic reconstructions and anticipate a broad applicability of this codification method.

Stories are not only painfully weaved by crafty writers in the solitude of their studios; they also have to be produced massively for non-player characters in the video game industry or tailored to particular tastes in personalized stories. However, the creation of fictional stories is a very complex task that usually implies a creative process where the author has to combine characters, conflicts and backstories to create an engaging narrative. This work describes a general methodology to generate cohesive and coherent backstories where desired archetypes (universally accepted literary symbols) can emerge in complex stochastic systems. This methodology supports the modeling and parametrization of the agents, the environment where they will live and the desired literary setting. The use of a Genetic Algorithm (GA) is proposed to establish the parameter configuration that will lead to backstories that best fit the setting. Information extracted from a simulation can then be used to create the literary work. To demonstrate the adequacy of the methodology, we perform an implementation using a specific multi-agent system and evaluate the results, testing with three different literary settings.

KANTS is a swarm intelligence clustering algorithm inspired by the behavior of social insects. It uses stigmergy as a strategy for clustering large datasets and, as a result, displays a typical behavior of complex systems: self-organization and global patterns emerging from the local interaction of simple units. This paper introduces a simplified version of KANTS and describes recent experiments with the algorithm in the context of a contemporary artistic and scientific trend called swarm art, a type of generative art in which swarm intelligence systems are used to create artwork or ornamental objects. KANTS is used here for generating color drawings from the input data that represent real-world phenomena, such as electroencephalogram sleep data. However, the main proposal of this paper is an art project based on well-known abstract paintings, from which the chromatic values are extracted and used as input. Colors and shapes are therefore reorganized by KANTS, which generates its own interpretation of the original artworks. The project won the 2012 Evolutionary Art, Design, and Creativity Competition.

In this work we address the problem of inventory and routing management in a retail chain. This involves the minimisation of two contradicting objectives, inventory holding costs and transportation costs, but which can be compounded in to a single one, the global costs. In previous work we addressed this using a single objective evolutionary algorithm but the duality inherent in the problem prompts us to consider a multiobjective approach; the aim is to determine what advantages each can bring. A number of experiments are carried out on several simulated and one real retail chain.

This paper presents a study on the effects of dynamic and partially connected 2-dimensional topologies on the performance of the particle swarm optimization (PSO). The swarm is positioned on 2-dimensional grids of nodes and the particles move through the nodes according to a simple rule. Meanwhile, the von Neumann neighborhood is used to decide which particles influence each individual. Structures with growing size are tested on a classical benchmark and compared to several configurations such as lbest, gbest and the standard von Neumann configuration. The results show that the partially connected grids with von Neumann neighborhood structure performs more consistently when compared to lbest, gbest and the standard von Neumann topology.

MasterMind is a puzzle in which a hidden string of symbols must be discovered by producing query strings which are compared with the hidden one; the result of this comparison (in terms of number of correct positions and colors) is fed back to the player that is trying to crack the code (codebreaker). Methods for solving this puzzle are usually compared in terms of the number of query strings (guesses) made and the total time needed to produce those strings. In this paper we focus on the latter by trying to find a combination of parameters that is, first, uniform and independent of the problem size, and second, adequate to find a fast solution that is, at the same time, good enough. The key to this combination of parameters will be the consistent set size, that is, the maximum number of combinations that are sought before being scored and played as a guess. Having found in previous papers that the consistent set size has an influence on speed, we will concentrate on small sizes and test them through two different scoring methods from literature: most parts and entropy to find out the influence of that parameter on the outcome and which method scales better. With this we try to find out which method and size yield the best results an are effectively able to? find solutions for sizes not approached so far in a reasonable time.

This paper presents a collaborative-interactive evolutionary algorithm (C-IEA) that evolves artistic animations and is executed on the web. The application is called Fireworks, since the animations that are produced are similar to an elaborate fireworks display. The system is built using the EvoSpace platform for distributed and asynchronous evolutionary algorithms. EvoSpace provides a central repository for the evolving population and remote clients, called EvoWorkers, that interact with the system to perform fitness evaluation using an interactive approach. The artistic animations are coded using the Processing programming language that facilitates rapid development of computer graphics applications for artists and graphic designers. The system promotes user collaboration and interaction by allowing many users to participate in population evaluation and because the system incorporates social networking. Initial results show that the proposed C-IEA can allow users to produce interesting artistic artifacts that incorporate preferences from several users, evolving dynamic animations that are unique within evolutionary art.

In this paper we present a distributed evolutionary algorithm that uses exclusively cloud services. This presents certain advantages, such as avoiding the acquisition of expensive resources, but at the same time presents the problem of choice between different services at different levels (infrastructure, platform, software) and, finally the actual scalability that can be achieved in a real distributed evolutionary algorithm. These issues are addressed by creating a pure-cloud version of EvoSpace, a pool-based evolutionary algorithm previously presented by the authors. EvoSpace is tested using the free tier of two services (one for the pool and other for the clients) and also the paying tier, and speedup is measured and its limits assessed. In general, this paper proves that a low-cost distributed evolutionary algorithm system can be created using cloud services that can be set up in very short time, but that major efficiency improvements can be obtained by switching to the non-free tier, giving another twist to the famous phrase “there is no free lunch”. We also show that using a pool-based algorithm allows to use cloud services more efficiently (and dynamically) than a static or synchronous service.

Solving the MasterMind puzzle, that is, finding out a hidden combination by using hints that tell you how close some strings are to that one is a combinatorial optimization problem that becomes increasingly difficult with string size and the number of symbols used in it. Since it does not have an exact solution, heuristic methods have been traditionally used to solve it; these methods scored each combination using a heuristic function that depends on comparing all possible solutions with each other. In this paper we first optimize the implementation of previous evolutionary methods used for the game of mastermind, obtaining up to a 40% speed improvement over them. Then we study the behavior of an entropy-based score, which has previously been used but not checked exhaustively and compared with previous solutions. The combination of these two strategies obtain solutions to the game of Mastermind that are competitive, and in some cases beat, the best solutions obtained so far. All data and programs have also been published under an open source license.

Pareto-based island model is a multi-colony distribution scheme recently presented for the resolution, by means of ant colony optimization algorithms, of bi-criteria problems. It yielded very promising results, but the model was implemented considering a unique Pareto-front-shaped unidirectional neighborhood migration topology, and a constant migration rate. In the present work two additional neighborhood topology schemes, and four different migration rates have been tested, considering the algorithm which obtained the best results in average in the model presentation article: MOACS (Multi-Objective Ant Colony System). Several experiments have been conducted, including statistical tests for better support the study. High values for the migration rate and the use of a bidirectional neighborhood migration topology yields the best results.

This paper compares the use of RGB and HSV histograms during the execution of an Evolutionary Algorithm. This algorithm generates abstract images that try to match the histograms of a target image. Three different fitness functions have been used to compare: the differences between the individual with the RGB histogram of the test image, the HSV histogram, and an average of the two histograms at the same time. Results show that the HSV fitness also increases the similarities of the RGB (and therefore, the average) more than the other two measures.

It is very important when search methods are being designed to know which parameters have the greatest influence on the behaviour and performance of the algorithm. To this end, algorithm parameters are commonly calibrated by means of either theoretic analysis or intensive experimentation. However, due to the importance of parameters and its effect on the results, finding appropriate parameter values should be carried out using robust tools to determine the way they operate and influence the results. When undertaking a detailed statistical analysis of the influence of each parameter, the designer should pay attention mostly to the parameters that are statistically significant. In this paper the ANOVA (ANalysis Of the VAriance) method is used to carry out an exhaustive analysis of an evolutionary algorithm method and the different parameters it requires. Following this idea, the significance and relative importance of the parameters regarding the obtained results, as well as suitable values for each of these, were obtained using ANOVA and post-hoc Tukey's Honestly Significant Difference tests on four well known function optimization problems. Through this statistical study we have verified the adequacy of parameter values available in the bibliography using parametric hypothesis tests.

This work presents the experimental results obtained with a distributed computing system created by mapping an evolutionary algorithm to the CouchDB object store. The framework decouples the population from the evolutionary algorithm and -through the API that CouchDB provides- allows the distributed and asynchronous operation of clients written in different programming languages. In this paper we present tests which prove that the novel algorithm design still performs as good as a canonical evolutionary algorithm and discover what are the main issues concerning it, what kind of speedups should we expect, and how all this affects the fundamental evolutionary algorithms concepts.

Cloud computing is lately becoming a part of the tool-set that the scientist uses to perform compute-intensive tasks. In particular, cloud storage is an easy and convenient way of storing files that will be accessible over the Internet, but can also be used for distributing those files for performing computation on them. In this paper we describe how such a service commercialized by Dropbox is used for pool-based evolutionary algorithms. A prototype system is described and its performance measured over deceptive combinatorial optimization problems using two different substrates: WiFi and wired, finding that, for some type of problems and using commodity hardware, cloud storage systems can profitably be used as a platform for distributed evolutionary algorithms; however, performance is influenced by the type of underlying network. After introducing the method in a previous paper, in this paper we focus on measuring this influence, finding that wired is faster than WiFi for any number of nodes. We have also performed an experiment with a few more computers to see whether speedup keeps up with the number of nodes.

Ant algorithms are known to return effective results in those problems that may be reduced to finding paths through a graph. However, this class of bio-inspired heuristics have raised the interest of the artistic community as well, namely of the artists that work on the blurred border between art and science. This paper describes an extension of an ant algorithm that, although has been designed as an edge detection tool and a model for collective perception, has also been used for creating artworks that were exhibited to a heterogeneous audience. The algorithm is a self-organized and stigmergic social insects' model that is able to evolve lines along the contours of an image, in a decentralized and local manner. The result is the emergence of global patterns called pheromone maps. These maps which were later named with the term pherographia are grayscale sketches of the original black-and-white image on top of which the model evolves. This work goes beyond grayscale images and addresses colored pherographia, by proposing several image transformation and border selection methods based on behavioral variations of the basic algorithm.

This paper describes an Evolutionary Algorithm for evolving the decision engine of a bot designed to play the Planet Wars game. This game, which has been chosen for the Google Artificial Intelligence Challenge in 2010, requires that the artificial player is able to deal with multiple objectives, while achieving a certain degree of adaptability in order to defeat different opponents in different scenarios. The decision engine of the bot is based on a set of rules that have been defined after an empirical study. Then, an Evolutionary Algorithm is used for tuning the set of constants, weights and probabilities that define the rules, and, therefore, the global behavior of the bot. The paper describes the Evolutionary Algorithm and the results attained by the decision engine when competing with other bots. The proposed bot defeated a baseline bot in most of the playing environments and obtained a ranking position in top-20% of the Google Artificial Intelligence competition.

The MasterMind puzzle is an interesting problem to be approached via evolutionary algorithms, since it is at the same time a constrained and a dynamic problem, and has eventually a single solution. In previous papers we have presented and evaluated different evolutionary algorithms to this game and shown how their behavior scales with size, looking mainly at the game-playing performance. In this paper we fine-tune the parameters of the evolutionary algorithms so that the worst case number of evaluations, and thus the average and median, are improved, resulting in a better solution in a more reliably predictable time.

Cloud computing, in general, is becoming part of the toolset that the scientist uses to perform compute-intensive tasks. In particular, cloud storage is an easy and convenient way of storing files that will be accessible over the Internet, but also a way of distributing those files and performing distributed computation using them. In this paper we describe how such a service commercialized by Dropbox is used for pool-based evolutionary algorithms. A prototype system is described and its peformance measured over a deceptive combinatorial optimization problem, finding that, for some type of problems and using commodity hardware, cloud storage systems can profitably be used as a platform for distributed evolutionary algorithms. Preliminary results show that Dropbox is indeed a viable alternative for execution of pool-based distributed evolutionary algorithms, showing a good scaling behavior with up to 4 computers.

The natural mate-selection behavior of preferring individuals which are somewhat (but not too much) different has been proved to increase the resistance to infection of the resulting offspring, and thus fitness. Inspired by these results we have investigated the improvement obtained from diversity induced by differences between individuals sent and received and the resident population in an island model, by comparing different migration policies, including our proposed multikulti methods, which choose the individuals that are going to be sent to other nodes based on the principle of multiculturality; the individual sent should be different enough to the target population, which will be represented through a proxy string (computed in several possible ways) in the emitting population. We have checked a set of policies following these principles on two discrete optimization problems of diverse difficulty for different sizes and number of nodes, and found that, in average or in median, multikulti policies outperform the usual policy of sending the best or a random individual; however, the size of this advantage changes with the number of nodes involved and the difficulty of the problem, tending to be greater as the number of nodes increases. The success of this kind of policies will be explained via the measurement of entropy as a representation of population diversity for the policies tested.

This paper proposes using the ANOVA (ANalysis Of the VAriance) method to carry out an exhaustive analysis of the simulated annealing (Sim-Ann) method and the different parameters it requires, such as those related to: the neighbourhood; the cooling scheme; the initial temperature; the number of times the cooling scheme is applied; and the number of times we search for best individual before the temperature is cooled. When undertaking a detailed statistical analysis of the influence of each parameter, the designer should pay attention mostly to the parameter presenting values that are statistically most significant. Following this idea, the significance and relative importance of the parameters with respect to the obtained results, as well as suitable values for each of these, were obtained using ANOVA on four well known function optimization problems.

In this paper we present an evolutionary approach to the joint management of inventory and routing in a retail chain. With this purpose, an ad-hoc evolutionary algorithm was designed which includes a non-standard individual representation and two mutation operators specific to this particular problem. Also, this work considers the inclusion of time-window constraints in the problem. We compare this approach to that of previous work, in which two levels were used: one for obtaining the delivery patterns that minimise the inventory costs and another to find the optimal routes in order to minimise transportation costs. The analysis performed allows us to conclude that the one-level approach obtains better results than the previously used two-level methodology in a shorter time.

In this paper we employ a steady state genetic algorithm to evolve different types of behaviour for bots in the Unreal Tournament 2004™computer game. For this purpose we define three fitness functions which are based on the number of enemies killed, the lifespan of the bot and a combination of both. Long run experiments were carried out, in which the evolved bots' behaviours outperform those of standard bots supplied by the game, particularly in those cases where the fitness involves a measure of the bot's lifespan. Also, there is an increase in the number of items collected, and the behaviours tend to become less aggressive, tending instead towards a more optimised combat style. Further “short run” experiments were carried out with a further type of fitness function defined, based on the number of items picked. In these cases the bots evolve performances towards the goal they have been aimed, with no other behaviours arising, except in the case of the multiple objective one. We conclude that in order to evolve interesting behaviours more complex fitness functions are needed, and not necessarily ones that directly include the goal we are aiming for.

FluidDB is a new structured storage system, available online for limited alpha test, which is designed to be able to easily store objects and relations among them (using tags). It is accessible through a simple REST interface, which is usually wrapped in a high-level language library. These features make them an ideal candidate for acting as the substrate of a persistent or pool based evolutionary algorithm, the Fluid Evolutionary Algorithm, presented for the first time in this paper. Our objective is to present a proof of concept and also to show how design decisions (about how and how often to use the pool, for instance) affect running time and algorithmic performance; we also show how FluidDB features positively affect algorithm design. These measures are mainly intended as a baseline measure, which can be improved on as FluidDB and fluid evolutionary algorithms (co-)evolve.

In this work we employ a steady state genetic algorithm to evolve bots' behaviors in the Unreal Tournament 2004 game. Our aim is to show whether interesting behaviors can be obtained with simple fitness functions. For this purpose we define four functions, measuring the number of enemies killed, the bot's lifespan, a combination of both and the number of items collected. The experiments show that incorporating a measure of the bot's lifespan in the fitness results in an optimal behavior in all aspects considered; further, the bots evolved this way outperform the standard bots supplied by the game. In addition, there is an increase in the number of items collected (even when this is not explicitly included in the fitness) and a tendency towards a more optimised combat style with less aggressive behaviors.

The Mastermind puzzle consists in finding out a secret combination by playing others in the same search space and using the hints obtained as a response (which reveal how close the played combination is to the secret one) to produce new combinations and eventually the secret one. Despite having been researched for a number of years, there are still several open issues, such as finding a strategy to select the next combination to play that is able to consistently obtain good results, at any problem size, and also doing it in as little time as possible. In this paper we cast this as a constrained optimization problem, introducing a new fitness function for evolutionary algorithms that takes that fact into account, and compare it to other solutions (exhaustive/heuristic and evolutionary), finding that it is able to obtain the consistently good solutions, and in as little as 30% less time than previously published evolutionary algorithms [2].

This work presents a feasible solution to the problem of book losses prediction from financial and general data in companies. The specific problem tackled in this work corresponds to a real dataset of Spanish companies. A Mutual Information-based criterion has been applied in order to reduce the initial set of variables, and a Support Vector Machine classifier has been designed to perform the prediction. The results show that the proposed approach obtains an important reduction of the number of variables needed to perform the prediction, improving the generalization capabilities of the model. The accuracy rates were above the 84% in the test set, much better than those obtained by other soft-computing algorithms (such as Genetic Programming, Self-Organizing Maps or Artificial Neural Networks) working with the same dataset and presented in previous works. The proposed approach shows to be promising and could be determinant in providing the experts with the right tools for the selection of the relevant factors and for the prediction in this difficult problem.

This paper reports the investigations and experimental procedures conducted for designing an automatic sleep classification tool basedconly in the features extracted with wavelets from EEG, EMG and EOG (electro encephalo-mio- and oculo-gram) signals, without any visual aid or context-based evaluation. Real data collected from infants was processed and classified by several traditional and bio-inspired heuristics. Preliminary results show that some methods are able to attain success rates close to 70% when compared to an expert neurologist. Although still not sufficient to implement a reliable sleep classifier, these are promising results that, together with an analysis via Self-Organizing Maps and ant-based clustering, may help to improve the feature extraction and contribute to a better representation of the different classes' characteristics.

This paper reports a comparison of several bloat control methods and also evaluates a recent proposal for limiting the size of the individuals: a genetic operator called prune and plant. The aim of this work is to test the adequacy of this method. Since a preliminary study of the method has already shown promising results, we have performed a thorough study in a set of benchmark problems aiming at demonstrating the utility of the new approach. Prune and plant has obtained results that maintain the quality of the final solutions in terms of fitness while achieving a substantial reduction of the mean tree size in all four problem domains considered. In addition, in one of these problem domains, prune and plant has demonstrated to be better in terms of fitness, size reduction, and time consumption than any of the other bloat control techniques under comparison. The experimental part of the study presents a comparison of performance in terms of phenotypic and genotypic diversity. This comparison study can provide the practitioner with some relevant clues as to which bloat control method is better suited to a particular problem and whether the advantage of a method does or does not derive from its influence on the genetic pool diversity.

Migration policies in distributed evolutionary algorithms are bound to have, as much as any other evolutionary operator, an impact on the overall performance. However, they have not been an active area of research until recently, and this research has concentrated on the migration rate. In this paper we compare different migration policies, including our proposed multikulti methods, which choose the individuals that are going to be sent to other nodes based on the principle of multiculturalism: the individual sent should be as different as possible to the receiving population (represented in several possible ways). We have checked this policy on two discrete optimization problems for different number of nodes, and found that, in average or in median, multikulti policies outperform others like sending the best or a random individual; however, their advantage changes with the number of nodes involved and the difficulty of the problem. The success of these kind of policies is explained via the measurement of entropies, which are known to have an impact in the performance of the evolutionary algorithm.

In this paper we compare different policies to select individuals to migrate in an island model. Our thesis is that choosing individuals in a way that exploits differences between populations can enhance diversity, and improve the system performance. This has lead us to propose a family of policies that we call multikulti, in which nodes exchange individuals different "enough" among them. In this paper we present a policy according to which the receiver node chooses the most different individual among the sample received from the sending node. This sample is randomly built but only using individuals with a fitness above a threshold. This threshold is previously established by the receiving node. We have tested our system in two problems previously used in the evaluation of parallel systems, presenting different degree of difficulty. The multikulti policy presented herein has been proved to be more robust than other usual migration policies, such as sending the best or a random individual.

Despite the intuition that the same population size is not needed throughout the run of an Evolutionary Algorithm (EA), most EAs use a fixed population size. This paper presents an empirical study on the possible benefits of a Simple Variable Population Sizing (SVPS) scheme on the performance of Genetic Algorithms (GAs). It consists in decreasing the population for a GA run following a predetermined schedule, configured by a speed and a severity parameter. The method uses as initial population size an estimation of the minimum size needed to supply enough building blocks, using a fixed-size selectorecombinative GA converging within some confidence interval toward good solutions for a particular problem. Following this methodology, a scalability analysis is conducted on deceptive, quasi-deceptive, and non-deceptive trap functions in order to assess whether SVPS-GA improves performances compared to a fixed-size GA under different problem instances and difficulty levels. Results show several combinations of speed-severity where SVPS-GA preserves the solution quality while improving performances, by reducing the number of evaluations needed for success.

This paper presents how it is possible to increase the Genetic Programming (GP) Computing Power (CP) for free, via Volunteer Computing (VC), using the well known framework BOINC plus a new ``virtualization'' layer which adds all the benefits from the virtualization paradigm. Two different experiments, employing a standard GP tool and a complex GP system, are performed --with distributed PCs over several cities-- to show the free achieved CP by means of VC, without the necessity of modifying or adapting the original GP source code. The methodology can be easily extended to Evolutionary Algorithms (EAs).

In this paper, we present a multiobjective ant colony optimization (MOACO) algorithm, called CHAC, designed to solve the problem of finding the path for a military unit that minimizes the cost in resources while maximizing safety. Unlike previous MOACO algorithms, CHAC uses a single colony and two different state transition rules: One that combines the heuristic and pheromone information of both objectives and another based on the dominance concept of multiobjective optimization problems. These rules have been evaluated in different scenarios (maps with different degrees of difficulty), outperforming a greedy algorithm (taken as baseline), and yielding a good military behavior in the tactical sense. In comparison, the combined rule is slightly better than the rule based on dominance. © 2009 Wiley Periodicals, Inc.

In many real world applications type I (false positive) and type II (false negative) errors have to be dealt with separately, which is a complex problem since an attempt to minimize one of them usually makes the other grow. In fact, a type of error can be more important than the other, and a trade-off that minimizes the most important error type must be reached. In the case of the bankruptcy prediction problem the error type II is of greater importance, being unable to identify that a company is at risk causes problems to creditors and slows down the taking of measures that may solve the problem. Despite the importance of type II errors, most bankruptcy prediction methods take into account only the global classification error. In this paper we propose and compare two methods to optimize both error types in classification: artificial neural networks and function trees ensembles created through multiobjective optimization. Since the multiobjective optimization process produces a set of equally optimal results (Pareto front) the classification of the test patterns in both cases is based on the non-dominated solutions acting as an ensemble. The experiments prove that, although the best classification rates are obtained using the artificial neural network, the multiobjective genetic programming model is able to generate comparable results in the form of an analytical function.

In this paper we present the Gossip-based Evolvable Agent Model (GossEvAg) within the context of parallel fine-grained Evolutionary Algorithms (EAs). It extends the Cellular Evolutionary Algorithm (CEA) definition with two novel features designed to work on Peer-to-Peer (P2P) networks: every individual is self-scheduled in a single thread and dynamically self-organizes its neighbourhood via newscasting, a gossip protocol. As a consequence of such multi-threading model, each Evolvable Agent (EvAg) updates asynchronously its state at random depending on the underlying platform scheduler. In order to assess the effects of asynchrony and the gossip protocol, we perform an experimental evaluation of the model for a set of discrete optimization problems. As a baseline for comparison we use two canonical genetic algorithms (GA): A steady-state GA (ssGA) and a generational GA (gGA). We also test two more topologies for the EvAg, a complete graph topology which allows panmixia and a Watts-Strogatz topology which has shown good theoretical and empirical results in related papers. We found that leaving the management of the EvAg to the underlying platform scheduler has an interesting emerging feature: the model is able to scale seamlessly in desktop computers without any effort from the practitioner. We measure how the algorithm speed scales by conducting the experiments in a Single and a Dual-Core Processor architectures.

In a connected world, spare CPU cycles are up for grabs, if you only make its obtention easy enough. In this paper we present a distributed evolutionary computation system that uses the computational capabilities of the ubiquituous web browser. Asynchronous Javascript and JSON (Javascript Object Notation, a serialization protocol) allows anybody with a web browser (that is, mostly everybody connected to the Internet) to participate in a genetic algorithm experiment with little effort, or none at all. Since, in this case, computing becomes a social activity and is inherently impredictable, in this paper we will explore the performance of this kind of virtual computer by solving simple problems such as the Royal Road function and analyzing how many machines and evaluations it yields. We will also examine possible performance bottlenecks and how to solve them, and, finally, issue some advice on how to set up this kind of experiments to maximize turnout and, thus, performance. The experiments show that we we can obtain high, and to a certain point, reliable performance from volunteer computing based on AJAJ, with speedups of up to several (averaged) machines.

This paper presents a statistical parameter analysis of the ant colony optimization algorithm that was implemented to solve the bi-criteria military path-finding problem. Three parameters have been studied using analysis of variance (ANOVA) in order to identify their influence in the results and the most suitable values for them: number of ants, number of iterations and exploration/exploitation factor. In addition, a mean analysis has been performed in order to complete the conclusions obtained. The study has yielded optimal values for the parameters under study, and some internal relationships between them have been identified.

The design of computer architectures is a very complex problem. The multiple parameters make the number of possible combinations extremely high.Many researchers have used simulation, although it is a slow solution since evaluating a single point of the search space can take hours. In this work we propose using evolutionary multilayer perceptron (MLP) to compute the performance of an architecture parameter settings. Instead of exploring the search space, simulating many configurations, our method randomly selects some architecture configurations; those are simulated to obtain their performance, and then an artificial neural network is trained to predict the remaining configurations performance. Results obtained show a high accuracy of the estimations using a simple method to select the configurations we have to simulate to optimize the MLP. In order to explore the search space, we have designed a genetic algorithm that uses the MLP as fitness function to find the niche where the best architecture configurations (those with higher performance) are located. Our models need only a small fraction of the design space, obtaining small errors and reducing required simulation by two orders of magnitude.

This paper investigates a new method for Genetic Algorithms' mutation rate control, based on the Sandpile Model: Sandpile Mutation. The Sandpile is a complex system operating at a critical state between chaos and order. This state is known as Self-Organized Criticality (SOC) and is characterized by displaying scale invariant behavior. In the precise case of the Sandpile Model, by randomly and continuously dropping "sand grains" on top of a two dimensional grid lattice, a power-law relationship between the frequency and size of sand "avalanches" is observed. Unlike previous off-line approaches, the Sandpile Mutation dynamics adapts during the run of the algorithm in a self-organized manner constrained by the fitness values progression. This way, the mutation intensity not only changes along the search process, but also depends on the convergence stage of the algorithm, thus increasing its adaptability to the problem context. The resulting system evolves a wide range of mutation rates during search, with large avalanches appearing occasionally. This particular behavior appears to be well suited for function optimization in dynamic environments, where large amounts of genetic novelty are regularly needed in order to track the moving extrema. Experimental results confirm these assumptions.

This paper introduces a procedure based on genetic programming to evolve XSLT programs (usually called stylesheets or logicsheets). XSLT is a general purpose, document-oriented functional language, generally used to transform XML documents or, in general, solve any problem that can be coded as an XML document. The proposed solution uses a tree representation for the stylesheets as well as diverse specific operators in order to obtain, in the studied cases and a reasonable time, a XSLT stylesheet that performs the transformation.

In this work we compare two soft-computing methods for producing models that are able to predict whether a company is going to have book losses: artificial neural networks (ANNs) and genetic programming (GP). In order to build prediction models that can be applied to an extensive number of practical cases, we need simple models which require a small amount of data. Kohonen's self-organizing map (SOM) is a non-supervised neural network that is usually used as a clustering tool. In our case a SOM has been used to reduce the dimensions of the prediction problem. Traditionally, ANNs have been considered able to produce better classifier structures than GP. In this work we merge the capability of GP for generating classification trees and the feature extraction abilities of SOM, obtaining a classification tool that beats the results yielded using an evolutionary ANN method.

The realization of a control unit can be done using a complex circuitry or microprogramming. The latter may be considered as an alternative method of implementation of machine instructions that can reduce the complexity and increase the flexibility of the control unit. The microcode efficiency and speed are of vital importance for the computer to execute machine instructions fast. This is a difficult task and it requires expert knowledge. It would be interesting and very helpful to have automated tools that, given a machine instruction description, could generate an efficient and correct microprogram. A good option is to use evolutionary computation techniques, which have proved been effective in the evolution of computer programs. In this paper, we intend to show how evolutionary computing techniques could be used to face this problem of generating efficient microprograms. We have developed a microarchitecture simulator of a real machine in order to evaluate an individual and to assign it the fitness value (to determine whether this candidate solution correctly implements the instruction machine). The proposed method is successful in generating correct solutions, not only for the machine code instruction set, but for new machine instructions not included in such set. We have shown that our approach can generate microprogramms to execute (to schedule microinstructions) the machine level instructions for a real machine. Moreover, this evolutive method could be applied to any microarchitecture just by changing the microinstruction set and pre-conditions of each machine instruction to guide evolution.

This paper reports a comparison of several bloat control methods and also evaluates a new proposal for limiting the size of the individuals: a genetic operator called prune and plant. The aim of this work is to prove the adequacy ofthis new method. Since a preliminary study of the methodhas already shown promising results, we have performed a thorough study in a set of benchmark problems aiming at demonstrating the utility of the new approach. Prune and plant has obtained results that maintain the quality ofthe final solutions in terms of fitness while achieving a substantial reduction of the mean tree size in all four problem domains considered. In addition, in one of these problem domains prune and plant has demonstrated to be better interms of fitness, size reduction and time consumption than any of the other bloat control techniques under comparison.

This paper investigates the behavior of the adaptive dissortative mating genetic algorithm (ADMGA) on dynamic problems and compares it with other genetic algorithms (GA). ADMGA is a non-random mating algorithm that selects parents according to their Hamming distance, via a self-adjustable threshold value. The resulting method, by keeping population diversity during the run, provides new means for GAs to deal with dynamic problems, which demand high diversity in order to track the optima. Tests conducted on combinatorial and trap functions indicate that ADMGA is more robust than traditional GAs and it is capable of outperforming a previously proposed dissortative scheme on a wide range of tests.

Radial basis function networks (RBFNs) have shown their capability to be used in classification problems, so that many data mining algorithms have been developed to configure RBFNs. These algorithms need to be given a suitable set of parameters for every problem they face, thus methods to automatically search the values of these parameters are required. This paper shows the robustness of a meta-algorithm developed to automatically establish the parameters needed to design RBFNs. Results show that this new method can be effectively used, not only to obtain good models, but also to find a stable set of parameters, available to be used on many different problems.

This paper presents an evolutionary algorithm for modeling the arrival dates of document streams, which is any time-stamped collection of documents, such as newscasts, e-mails, scientific journals archives and weblog postings. The goal is to find a frequency curve that fits the data circumventing the unavoidable noise. Classical dynamic programming algorithms are limited by memory and efficiency requirements, which can be a problem when dealing with long streams. This suggests to explore alternative search methods which although do not guarantee optimality, are far more efficient. Experiments have shown that the designed evolutionary algorithm is able to reach high quality solutions in a short time. We have also explored different approaches to infer whether new arrivals increase or decrease interest in the topic the document stream is about. In particular, we present a variant of the evolutionary algorithm, which is able to very quickly fit a stream extended with new data, by taking advantage of the fit obtained for the original substream. These mechanisms can be used for real time detection of changes in the trend of interest in a topic, an important application of this kind of models.

EVITA, standing for Evolutionary Inventory and TransportationAlgorithm, aims to be a commercial tool to addressthe problem of minimising both the transport and inventorycosts of a retail chain that is supplied from a centralwarehouse. In this paper we study different issues involvedin finding the appropriate settings for EVITA, so that itcan be employed by a non-expert user over wide range ofproblems.The aim is not to define a new algorithm for resolutionof the ITP, but to determine whether it is possible to finda set of input parameters that can provide good results ona wide range of problem configurations, hence eliminatingthe need for user adjustment once the tool is employed in acommercial setting.We focus on the influence of three parameters: the populationsize, the tournament size and the mutation probability.After extensive experimentation and statistical analysis weare able to find a good configuration for the three factors.

No abstract available.

Speech segmentation is one of the problems in the speech processing area. The main techniques that attempt to solve it are manual segmentation and hidden Markov model alignment. In this work a new technique based on an evolutionary algorithm that permits to segment the speech without a previous training process is presented.

Interest in hybrid methods that combine artificial neural networks and evolutionary algorithms has grown in the last few years, due to their robustness and ability to design networks by setting initial weight values, by searching the architecture and the learning rule and parameters. This paper presents an exhaustive analysis of the G-Prop method, and the different parameters the method requires (population size, selection rate, initial weight range, number of training epochs, etc.) are determined. The paper also the discusses the influence of the application of genetic operators on the precision (classification ability or error) and network size in classification problems. The significance and relative importance of the parameters with respect to the results obtained, as well as suitable values for each, were obtained using the ANOVA (analysis of the variance). Experiments show the significance of parameters concerning the neural network and learning in the hybrid methods. The parameters found using this method were used to compare the G-Prop method both to itself with other parameter settings, and to other published methods.

The Web newspaper pagination problem consists of optimizing the layout of a set of articles extracted from several Web newspapers and sending it to the user as the result of a previous query. This layout should be organized in columns, as in real newspapers, and should be adapted to the client Web browser configuration in real time. This paper presents an approach to the problem based on simulated annealing (SA) that solves the problem on-line, adapts itself to the client's computer configuration, and supports articles with different widths.

Most genetic algorithm (GA) users adjust the main parameters of the design of a GA (crossover and mutation probability, population size, number of generations, crossover, mutation, and selection operators) manually. Nevertheless, when GA applications are being developed it is very important to know which parameters have the greatest influence on the behavior and performance of a GA. The purpose of this study was to analyze the dynamics of GAs when confronted with modifications to the principal parameters that define them, taking into account the two main characteristics of GAs; their capacity for exploration and exploitation. Therefore, the dynamics of GAs have been analyzed from two viewpoints. The first is to study the best solution found by the system, i.e., to observe its capacity to obtain a local or global optimum. The second viewpoint is the diversity within the population of GAs; to examine this, the average fitness was calculated. The relevancy and relative importance of the parameters involved in GA design are investigated by using a powerful statistical tool, the analysis of the variance (ANOVA).

This paper describes a project funded by the European Commission which seeks to provide the technology and software infrastructure necessary to support the next generation of evolving infohabitants in a way that makes that infrastructure universal, open and scalable. The Distributed Resource Evolutionary Algorithm Machine (DREAM) will use existing hardware infrastructure in a more efficient manner, by utilising otherwise unused CPU time. It will allow infohabitants to co-operate, communicate, negotiate and trade; and emergent behaviour is expected to result. It is expected that there will be an emergent economy that results from the provision and use of CPU cycles by infohabitants and their owners. The DREAM infrastructure will be evaluated with new work on distributed data mining, distributed scheduling and the modelling of economic and social behaviour.

A general problem in model selection is to obtain the right parameters that make a model fit observed data. For a multilayer perceptron (MLP) trained with backpropagation (BP), this means finding the right hidden layer size, appropriate initial weights and learning parameters. The paper proposes a method (G-Prop-II) that attempts to solve that problem by combining a genetic algorithm (GA) and BP to train MLPs with a single hidden layer. The GA selects the initial weights and the learning rate of the network, and changes the number of neurons in the hidden layer through the application of specific genetic operators. G-Prop-II combines the advantages of the global search performed by the GA over the MLP parameter space and the local search of the BP algorithm. The application of the G-Prop-II algorithm to several real world and benchmark problems shows that MLPs evolved using G-Prop-II are smaller and achieve a higher level of generalization than other perceptron training algorithms, such as QuickPropagation or RPROP, and other evolutive algorithms, such as G-LVQ. It also shows some improvement over previous versions of the algorithm.