Finding the right balance between exploration and exploitation is a crucial challenge in evolutionary algorithms (EAs), as it strongly impacts solution quality and the ability to escape local optima. While diversity-preserving mechanisms are commonly employed, high population diversity does not necessarily translate into improved performance across problem landscapes. Recent empirical evidence suggests that effective EA optimization largely depends on predominantly exploitative search dynamics, with exploration being most beneficial during the early iterations. Nevertheless, sporadic exploratory bursts may help reintroduce diversity and enhance solution quality. In this study, inspired by dropout regularization in neural networks (NNs), we introduce three mechanisms—and their variants— that stochastically deactivate entire individuals or subsets of decision variables with the expectation of enhancing exploration. We assess the effectiveness of dropout in genetic algorithms (GAs) on twelve single-objective and five multi-objective benchmark functions. We empirically demonstrate that per-variable stochastic deactivation, termed Individual Dropout (IDrop), significantly outperforms a standard GA on single-objective problems with exploitable global structure and improves convergence toward the theoretical Pareto front in multi-objective settings. We also show that, while dropout alone is not exempt from the curse of dimensionality, combinations of dropout mechanisms can partially alleviate this limitation in most of the tested functions.

Abstract Feature selection is one of the most important concepts in data mining when dimensionality reduction is needed. The performance measures of feature selection encompass predictive accuracy and result comprehensibility. Consistency‐based methods are a significant category of feature selection research that substantially improves the comprehensibility of the result using the parsimony principle. In this work, the biobjective version of the algorithm logical analysis of inconsistent data is applied to large volumes of data. In order to deal with hundreds of thousands of attributes, heuristic decomposition uses parallel processing to solve a set covering problem and a cross‐validation technique. The biobjective solutions contain the number of reduced features and the accuracy. The algorithm is applied to omics datasets with genome‐like characteristics of patients with rare diseases.

Automatic dimensionality reduction in text classification requires large training data sets due to the high dimensionality of the native feature space. However, in several real world multi-label problems, such as highly demanding decision-making scenarios, to manually classify and select features in large document sets is usually unfeasible even by specialist teams. This paper presents CrowdFS a first approach on using collective intelligence techniques to select label specific relevant features from a large document set. An experiment in the context of competitive intelligence for a multinational energy company showed CrowdFS producing better results than an automatic state of the art technique.

A long-standing goal in artificial intelligence and robotics is synthesising agents that can effectively learn and adapt throughout their lifetime. One open-ended approach to behaviour learning in autonomous robots is online evolution , which is part of the evolutionary robotics field of research. In online evolution approaches, an evolutionary algorithm is executed on the robots during task execution, which enables continuous optimisation and adaptation of behaviour. Despite the potential for automatic behaviour learning, online evolution has not been widely adopted because it often requires several hours or days to synthesise solutions to a given task. In this respect, research in the field has failed to develop a prevalent algorithm able to effectively synthesise solutions to a large number of different tasks in a timely manner. Rather than focusing on a single algorithm, we argue for more general mechanisms that can combine the benefits of different algorithms to increase the performance of online evolution of robot controllers. We conduct a comprehensive assessment of a novel approach called online hyper-evolution (OHE). Robots executing OHE use the different sources of feedback information traditionally associated with controller evaluation to find effective evolutionary algorithms during task execution. First, we study two approaches: OHE-fitness, which uses the fitness score of controllers as the criterion to select promising algorithms over time, and OHE-diversity, which relies on the behavioural diversity of controllers for algorithm selection. We then propose a novel class of techniques called OHE-hybrid, which combine diversity and fitness to search for suitable algorithms. In addition to their effectiveness at selecting suitable algorithms, the different OHE approaches are evaluated for their ability to construct algorithms by controlling which algorithmic components should be employed for controller generation (e.g., mutation, crossover, among others), an unprecedented approach in evolutionary robotics. Results show that OHE (i) facilitates the evolution of controllers with high performance, (ii) can increase effectiveness at different stages of evolution by combining the benefits of multiple algorithms over time, and (iii) can be effectively applied to construct new algorithms during task execution. Overall, our study shows that OHE is a powerful new paradigm that allows robots to improve their learning process as they operate in the task environment.

One of the long-term goals in evolutionary robotics is to be able to automatically synthesize controllers for real autonomous robots based only on a task specification. While a number of studies have shown the applicability of evolutionary robotics techniques for the synthesis of behavioral control, researchers have consistently been faced with a number of issues preventing the widespread adoption of evolutionary robotics for engineering purposes. In this article, we review and discuss the open issues in evolutionary robotics. First, we analyze the benefits and challenges of simulation-based evolution and subsequent deployment of controllers versus evolution on real robotic hardware. Second, we discuss specific evolutionary computation issues that have plagued evolutionary robotics: (1) the bootstrap problem, (2) deception, and (3) the role of genomic encoding and genotype-phenotype mapping in the evolution of controllers for complex tasks. Finally, we address the absence of standard research practices in the field. We also discuss promising avenues of research. Our underlying motivation is the reduction of the current gap between evolutionary robotics and mainstream robotics, and the establishment of evolutionary robotics as a canonical approach for the engineering of autonomous robots.

Symbiogenetic MuGA (SMuGA) is a co-evolutionary model exploiting the concept of symbiosis over the Multiset Genetic Algorithm (MuGA). It evolves two species: hosts that represent a solution to the problem, and parasites that represent part-solutions. SMuGA has been proved valuable in the optimization of a variety of deceptive functions. However its performance decreased in large scale difficult problems. This paper presents a new version of SMuGA with improvements centered on the evolution process of the parasites. The most significant advance is provided by using a diversity measure to modulate the evolution of parasites. The algorithm is tested with very good results in deceptive functions with up to 1024 bits. The paper is concluded with an analysis of the advantages and limitations of the approach, and with perspectives for future developments.

Online evolution gives robots the capacity to learn new tasks and to adapt to changing environmental conditions during task execution. Previous approaches to online evolution of neural controllers are typically limited to the optimisation of weights in networks with a prespecified, fixed topology. In this article, we propose a novel approach to online learning in groups of autonomous robots called odNEAT. odNEAT is a distributed and decentralised neuroevolution algorithm that evolves both weights and network topology. We demonstrate odNEAT in three multirobot tasks: aggregation, integrated navigation and obstacle avoidance, and phototaxis. Results show that odNEAT approximates the performance of rtNEAT, an efficient centralised method, and outperforms IM-(μ + 1), a decentralised neuroevolution algorithm. Compared with rtNEAT and IM-(μ + 1), odNEAT's evolutionary dynamics lead to the synthesis of less complex neural controllers with superior generalisation capabilities. We show that robots executing odNEAT can display a high degree of fault tolerance as they are able to adapt and learn new behaviours in the presence of faults. We conclude with a series of ablation studies to analyse the impact of each algorithmic component on performance.

In this paper, we discuss how the combination of evolutionary techniques and engineering-oriented approaches is an effective methodology for leveraging the potential of evolutionary robotics (ER) in the synthesis of behavioural control. We argue that such combination can eliminate the issues that have prevented ER from scaling to complex real-world tasks,namely: (i) the bootstrap problem, (ii) deception, (iii) the reality gap effect, and (iv) the prohibitive amount of time necessary to evolve controllers directly in real robotic hardware. We present recent studies carried out in our research group involving real-robot and simulation-based experiments. We provide examples of how the synergistic effects of evolution and engineering overcome each other’s limitations and significantly extend their respective capabilities, thereby opening a new path in the design of robot controllers.

MuGA is an evolutionary algorithm (EA) that represents populations as multisets, instead of the conventional collection. Such representation can be explored to adapt genetic operators in order to increase performance in difficult problems. In this paper we present an adaptation of the mutation operator, multiset wave mutation (MWM), that explores the multiset representation to apply different mutation ratios to the same chromosome, and an adaptation of the replacement operator, multiset decimation replacement (MDR) that preserves multiset representation in the main population and helps MuGA to solve hard deceptive problems. Results obtained in different deceptive functions show that pairing both operators is a robust approach with a high success ratio in most of the problems.

This paper presents ARES-III, a multi-purpose service robot for robust operation in all-terrain outdoor environments. Currently in pre-production phase, ARES-III is aimed to fulfil the requirements of a robotic platform that is able to support the development of real world applications in surveillance, agriculture, environmental monitoring, and other related domains. These demanding scenarios motivate a design focused on the reliability of the mechanical platform, the scalability of the control system, and the flexibility of its self-diagnosis and error recovery mechanisms. These are key features of ARES-III often disregarded in current commercial and research platforms. First, a comprehensive set of field trials demonstrated the ability of the ARES-III chassis, made of durable materials and with no-slip quasi-omnidirectional kinematic characteristics, to perform robustly in rough terrain. Second, supported by a control system fully compliant with the wide spread Robot Operating System (ROS), the scalability of ARES-III is enforced. Finally, the integration of active self-diagnosis and error recovery mechanisms in ARES-III control system fosters long lasting operation.

Recent approaches to appearance-based robot localization and mapping have relied on a topological representation of the environment and loop closing procedures to build correct maps. Because these solutions do not separate the problem of localization in mapped areas from the exploration of new locations, they require that loop-closing detection is performed continuously, inducing an extra computational burden on the robot. In this paper, we introduce a modified version of the Markov localization algorithm which avoids this condition by employing procedures that detect when novel areas are being explored or already mapped areas are revisited.

This paper presents a saliency-based solution to boost trail detection. The proposed model builds on the empirical observation that trails are usually conspicuous structures in natural environments. This hypothesis is confirmed by the experimental results, where a strong positive correlation between trail location and visual saliency has been observed. These results are due in part to the proposed extensions to a well known visual saliency computational model. This paper goes further and shows that, with a proper analysis of the saliency information alone, the ambiguity regarding both trail's position and approximate skeleton is reduced to three hypotheses in 98% of the tested natural images. This analysis is performed by a set of agents inhabiting the saliency and feature specific intermediate maps. These agents' behaviours exploit implicit, top-down knowledge about the object being sought in an active way. With the proposed model, computationally demanding accurate trail detectors are able to focus their activity in a fraction of the input image, thus promoting robustness and real-time performance.

Stereo-based off-road obstacle detection is a complex and still open problem. The challenges are in great extent related to computational cost and noise level. Previous work has shown that visual saliency and voting mechanisms are extremely effective in tackling these issues. This paper proposes a set of extensions to these mechanisms, to further improve the detector's speed-accuracy trade-off as well as its robustness. The observed enhancements are due in part to the adaptive way saliency is accounted for by the detector during the image scanning procedure. Additionally, detector's positive results are in turn used to boost the saliency map itself, thus reinforcing the analysis of relevant regions of the image. To enable detection in highly roughed terrain, the detector's invariance to the robot's posture is also enhanced. Experimental results show that, with the extended detector, denominated of ESalOD, higher levels of robustness, accuracy and computational efficiency are attained.

This paper proposes a model for trail detection that builds upon the observation that trails are salient structures in the robot's visual field. Due to the complexity of natural environments, the straightforward application of bottom-up visual saliency models is not sufficiently robust to predict the location of trails. As for other detection tasks, robustness can be increased by modulating the saliency computation with top-down knowledge about which pixel-wise visual features (e.g., colour) are the most representative of the object being sought. This paper proposes the use of the object's overall layout instead, as it is a more stable and predictable feature in the case of natural trails. This novel component of top-down knowledge is specified in terms of perception-action rules, which control the behaviour of simple agents performing as a swarm to compute the saliency map of the input image. For the purpose of multi-frame evidence accumulation about the trail location, a motion compensated dynamic neural field is used. Experimental results on a large data-set reveal the ability of the model to produce a success rate of 91% at 20Hz. The model shows to be robust in situations where previous trail detectors would fail, such as when the trail does not emerge from the lower part of the image or when it is considerably interrupted.

Inspection of an environment is often limited to a single video feed, provenient from a camera assembled on a mobile robot. This is known as the keyhole effect. In addition to this, the operator must often perform several tasks at the same time and specially in cluttered environments it can be hard to navigate the robot and at the same time keep the viewpoint of a camera aligned with a specific point of interest. This paper tackles this problem by proposing two mechanisms for assisting in teleoperation by positioning a pan-tilt camera. The goal consists in improving the teleoperatorpsilas situation awareness. Field trials with a pan-tilt-zoom (PTZ) camera mounted on an all-terrain robot validate the proposed mechanisms.

Human judgement is an integral part of the teleoperation process, heavily influenced by the rate in which commands are given to and the telemetry data is received from the robot. In this paper a multi-agent system for robot teleoperation is proposed. Several entities (e.g. robot or human) can easily be added and removed. In addition, this approach fosters a shared notion of reality to every entity present in the network, through a mechanism similar to the distributed blackboard architecture. This method provided an exchange of information regarding the teleoperation scenario, i.e. perceptual clues. This promoted a study on how perceptual clues influenced the operators’ judgement and performance. The experimental results in a physical suggest that the system is able to guarantee a close interaction between users and robots.

The unstructured nature of cross-country environments poses a great deal of challenges to obstacle detection algorithms. This paper presents an algorithm able to perform obstacle detection with small computational cycles, and extremely robust against outliers. To reduce the computational cost, foveation, i.e. space-variant image resolution adaptation, is employed. To reduce sensitivity to outliers, two voting filters are presented. Experimental results on two different stereo vision heads show that the proposed algorithm is able to perform at 10 Hz. Even in the presence of a low signal-to-noise ratio, caused by the use of a partially damaged lens, the algorithm is able to distinguish obstacles from the background.

As information infrastructure move towards open systems where agents come and go, new facilities are required so that these agents can take advantage of each other's functionalities. We need agent systems that can provide to newcomer agents a place and the right agent to interact with. Such functionality must cope with high rate of agent entrance, with high load of agents, with vanishing agents or nodes in the agent system. Given these requirements, agents are constantly facing a problem of deciding where to go and with whom to work with. These two decisions, pertaining to mobility and interaction, have been singled out as fundamental for every agent system. We present an algorithm targeted at these two decisions while it fulfils the aforementioned requirements.

The adoption of agents as utile companions faces the problem of conciliating the development of complex and intelligent functionalities with the requirements of autonomy mobility and adaptability. Our main focus will be on the agents adaptability. A hybrid agent architecture approach is proposed where a static component, which resides at the user's host and includes most of the intelligence and decision support capabilities, is complemented by a mobile component that is aimed at interacting with other agents. Some adaptation strategies, based on classical and fuzzy methodologies, are also discussed using as background scenario a trading market competitive environment with buyer and seller agents interacting in it.

This work attempts to shed light on the fundamental concepts behind the stability of multi-agent systems. We view the system as a discrete time Markov chain with a potentially unknown transitional probability distribution. The system will be considered to be stable when its state has converged to an equilibrium distribution. Faced with the non-trivial task of establishing the convergence to such a distribution, we propose a hypothesis testing approach according to which we test whether the convergence of a particular system metric has occurred. We describe some artificial multi-agent ecosystems that were developed and we present results based on these systems which confirm that this approach qualitatively agrees with our intuition.

In a trading scenario agents interact with each other, selling and buying resources. In order to control the behavior of the trading scenario, the interactions must be coordinated. We present a brief discussion of communication types and coordination models applicable in multi-agent systems. We find a programmable tuple space more appropriate to manage and rule the interactions between the trading agents. We discuss the advantages of a trading agent model that deals with the trading strategy, concentrating on what to buy or sell. This relieves the agent from the task of coordinating the negotiations and their revoking or acceptances. This is the task of the programmable tuple space.

A fuzzy behavior based decision control architecture is introduced. Each behavior is composed by several fuzzy actions-a concept that constitutes the behavior building block, allowing the implementation of a single aspect of the desired competence. A fuzzy action itself is composed by action (fuzzy controller) and activity (fuzzy predicate) producing modules. The behavior performance is also dependent on the (simulated) available energy. The arbitration process is present at the levels of action and behavior. This architecture is tested on the Khepera robot, both in simulation and in reality. The results of the performed experiments are presented, encouraging the architecture use in long range autonomous vehicles.

An architecture for the control of an autonomous vehicle for underwater operation is described, focusing on the important characteristics for autonomy. This approach considers a layered structure of interacting and concurrent modules, called behaviours, each one trying to perform its own independent activity, altogether producing an 'intelligent' global behaviour for the vehicle. The behaviours are kept simple, and their short term memory and combined performance suggest a reactive system. A layer for avoiding traps and deadlock situations is proposed, as a means of producing some sort of planning, which is not typical of reactive architectures. Explanations of natural systems are analyzed and exploited as guides for the design of autonomous underwater vehicles. The world model of the AUV is defined and divided in two major components. Its integration in the control architecture is described. The interface with the user is an important issue in this type of vehicles and therefore its support, namely at the world model is discussed. A simple case study in the end tries to illustrate the presented approach.>