Francisco Bellas

Real-valued multimodal fitness landscape characterization for evolution

This paper deals with the characterization of the fitness landscape of multimodal functions and how it can be used to choose the most appropriate evolutionary algorithm for a given problem. An algorithm that obtains a general description of real valued multimodal fitness landscapes in terms of the relative number of optima, their sparseness, the size of their attraction basins and the evolution of this size when moving away from the global optimum is presented and used to characterize a set of well-known multimodal benchmark functions. To illustrate the relevance of the information obtained and its relationship to the performance of evolutionary algorithms over different fitness landscapes, two evolutionary algorithms, Differential Evolution and Covariance Matrix Adaptation, are compared over the same benchmark set showing their behavior depending on the multimodal features of each landscape.

Considerations in the application of evolution to the generation of robot controllers

Abstract This paper is concerned with different aspects of the use of evolution for the successful generation of real robot Artificial Neural Network (ANN) controllers. Several parameters of an evolutionary/genetic algorithm (GA) and the way they influence the evolution of ANN behavioral controllers for real robots have been contemplated. These parameters include the way the initial populations are distributed, how the individuals are evaluated, the implementation of race schemes, etc. A batch of experiments on the evolution of three types of behaviors with different population sizes have been carried out in order to ascertain their effect on the evolution of the controllers and their validity in real implementations. The results provide a guide to the design of evolutionary algorithms for generating ANN based robot controllers, especially when, due to computational constraints, the populations to be used are small with respect to the complexity of the problem to be solved. The problem of transferring the controllers evolved in simulated environments to the real systems operating in real environments are also considered and we present results of this transference to reality with a robot which has few and extremely noisy sensors.

Specialization analysis of embodied evolution for robotic collective tasks

Abstract The objective of this work is to analyze embodied evolution based algorithms in coordinated multi-robot tasks that require specialization. This type of algorithm performs a Darwinian open-ended evolution where the individuals that make up the population are embodied in the physical robots and situated in an environment. The robots interact autonomously in an asynchronous fashion, leading to a complex dynamic system in continuous evolution with dependencies among parameters that make theoretical studies of specialization quite difficult in real cases. Consequently, the aim here is to perform a theoretical analysis of this type of embodied evolution based algorithms, establishing a set of canonical parameters that define their operation. A generic algorithm of this type is designed that allows us to formally study the relevance of the canonical parameters. In this paper this study concentrates on specialization for the construction of heterogeneous robotic teams. The conclusions obtained in the theoretical framework are confirmed in a real multi-robot collective gathering task using one of the many real embodied evolution based algorithms and showing that two canonical parameters are the most relevant in terms of specialization for this type of algorithms. Some insights into how to adjust these canonical parameters in a real problem are provided.

EDHMoR: Evolutionary designer of heterogeneous modular robots

This paper is devoted to the problem of automatically designing feasible and manufacturable robots made up of heterogeneous modules. Specifically, the coevolution of morphology and control in robots is analyzed and a particular strategy to address this problem is contemplated. To this end, the main issues of this approach such as encoding, evaluation or transfer to reality are studied through the use of heterogeneous modular structures with distributed control. We also propose a constructive evolutionary algorithm based on tree-like representations of the morphology that can intrinsically provide for a type of generative evolutionary approach. The algorithm introduces some new elements to smooth the search space and make finding solutions much easier. The evaluation of the individuals is carried out in simulations and then transferred to real robots assembled from the modules considered. To this end, the extension of the principles proposed by classical authors in traditional evolutionary robotics to brain-body evolution regarding how simulations should be set up so that robust behaviors that can be transferred to reality are obtained is considered here. All these issues are analyzed by means of an evolutionary design system called EDHMoR (Evolutionary Designer of Heterogeneous Modular Robots) that contains all the elements involved in this process. To show practical evidences of the conclusions that have been extracted with this work, two benchmark problems in modular robotics are considered and EDHMoR is tested over them. The first one is focused on solving a linear robot motion mission and the second one on a static task of the robot that does not require displacements.

An Adaptive Approach for the Progressive Integration of Spatial and Spectral Features When Training Ground-Based Hyperspectral Imaging Classifiers

The use of hyperspectrometers as analytical tools for determining surface material properties in ground-based applications introduces the need of integrating spatial and spectral hyperspectral cube components. A neural-network-based approach is presented in this paper with the aim of automatically adapting to the spatiospectral characteristics of samples in a problem domain so that the most efficient classification can be obtained. Its main application would be in inspection and quality control tasks. The system core is an Artificial Neural Network-based hyperspectral processing unit able to perform the online classification of the material based on the spatiospectral patterns provided by a set of pixels. A training adviser is implemented to automate the determination of the minimum spatial window size, as well as the optimum spectrospatial feature set leading to the desired classification in terms of the available ground truth. Several tests have been carried out on synthetic and real data sets. In particular, the proposed approach is used to discriminate samples of synthetic and real materials, where the pixel resolution implies that a material is characterized by spectral patterns of combinations of pixels.

Hyperspectral image segmentation through evolved cellular automata

Segmenting multidimensional images, in particular hyperspectral images, is still an open subject. Two are the most important issues in this field. On one hand, most methods do not preserve the multidimensional character of the signals throughout the segmentation process. They usually perform an early projection of the hyperspectral information to a two dimensional representation with the consequent loss of the large amount of spectral information these images provide. On the other hand, there is usually very little and dubious ground truth available, making it very hard to train and tune appropriate segmentation and classification strategies. This paper describes an approach to the problem of segmenting and classifying regions in multidimensional images that performs a joint two-step process. The first step is based on the application of cellular automata (CA) and their emergent behavior over the hyperspectral cube in order to produce homogeneous regions. The second step employs a more traditional SVM in order to provide labels for these regions to classify them. The use of cellular automata for segmentation in hyperspectral images is not new, but most approaches to this problem involve hand designing the rules for the automata and, in general, average out the spectral information present. The main contribution of this paper is the study of the application of evolutionary methods to produce the CA rule sets that result in the best possible segmentation properties under different circumstances without resorting to any form of projection until the information is presented to the user. In addition, we show that the evolution process we propose to obtain the rules can be carried out over RGB images and then the resulting automata can be used to process multidimensional hyperspectral images successfully, thus avoiding the problem of lack of appropriately labeled ground truth images. The procedure has been tested over synthetic and real hyperspectral images and the results are very competitive.

An evolution friendly modular architecture to produce feasible robots

This paper proposes the use of a modular robotic architecture in order to produce feasible robots through evolution. To this end, the main requirements the architecture must fulfill are analyzed and a top-down methodology is employed to obtain the different types of modules that make it up. Specifically, the problem of how to increase the evolvability or evolution friendliness of the system is addressed by considering a heterogeneous modular architecture with a large number of connection faces per module. Afterwards, a prototypical implementation of these modules with the required features is described and different experiments provide an indication of how versatile the architecture is for evolving robot morphologies and control for specific tasks and how easy it is to build them. We present a modular architecture to produce feasible robots through evolution.The architecture is based on a set of a heterogeneous modules.The modules contain a large number of connection faces per module.The design and the implementation of prototype modules is described in detail.Different experiments show its potential for evolving robot morphologies and control.

A comparison of gaussian based ANNs for the classification of multidimensional hyperspectral signals

This paper is concerned with the comparison of three types of Gaussian based Artificial Neural Networks in the very high dimensionality classification problems found in hyperspectral signal processing. In particular, they have been compared for the spectral unmixing problem given the fact that the requirements for this type of classification are very different from other realms in two aspects: there are usually very few training samples leading to networks that are very easily overtrained, and these samples are not usually representative in terms of sampling the whole input-output space. The networks selected for comparison go from the classical Radial Basis Function (RBF) network to the more complex Gaussian Synapse Based Network (GSBN) considering an intermediate type, the Radial Basis Function with Multiple Deviation (RBFMD). The comparisons were carried out when processing a benchmark set of synthetic hyperspectral images containing mixtures of spectra from materials found in the US Geological Service database.

Real-time optimization of dynamic problems through distributed Embodied Evolution

Several engineering optimization problems like routing, freight transportation, exploration, or layout design in their more complex and realistic versions present a series of characteristics that make them very difficult to solve. Among these we find the absence of centralized updated information about all the variables, due to the spread out nature of the problems or lack of appropriate communications, or the dynamism of real-time operation. In fact, most optimization approaches assume that the problems they address are static, meaning that there is an optimal solution that does not change in time, but this is not always the case and there are problems that require following an optimum that changes in time. Distributed population-based techniques, such as swarms, have provided promising results in this context. They obtain a solution through the concurrent behavior of several adequately constructed processing elements. However, constructing these swarms is not straightforward, and most approaches have just mimicked swarm behaviors found in nature, adapting them to particular problems. The objective of this work is to study the application of a novel evolutionary paradigm, distributed Embodied Evolution (dEE), to obtain heterogeneous swarms that solve a set of realistic problems. In particular, we address here non-separable dynamic fitness landscapes, where interdependences between individuals imply that the contribution provided by one of them to the whole depends on the behavior of the others. This study is carried out applying a canonical version of dEE, which has been developed to generalize the main features of this type of evolutionary paradigm. We analyze the canonical dEE response in a series of scenarios of increasing complexity related to two highly representative dynamic engineering problems: a Dynamic Fleet Size and Mix Vehicle Routing Problem with Time Windows (DFSMVRPTW) and a collective surveillance task with realistic location degradation.

Towards an evolutionary design of modular robots for industry

We are interested in the next generation of industrial robots, those that are able to operate in dynamic and unstructured environments and, consequently, that are able to adapt to changing circumstances or to work on different tasks in an autonomous way. In this sense, multirobot systems and, in particular, modular systems present several features like scalability, fault tolerance, low maintenance or reconfiguration capabilities that make them highly suitable for this kind of environments. The work presented here is concerned with the problem of automatically obtaining the morphology and control structure for this type of modular systems. In this line, we present the first results produced using a newly designed constructive evolutionary approach that takes into account the extreme difficulty of the tremendously deceptive and uninformative search space this type of applications are faced with. As an example, the algorithm is used to design the morphology and the distributed control parameters for a typical benchmark problem, that of moving as far as possible in a straight line, for a heterogeneous modular robotic system developed by our group.