This article introduces a new version of the car-pooling problem (CPP). This involves defining rendezvous or meeting point in such a way that the travel times of the users are fair, this problem shares similarities with the problem of finding the optimal meeting point (OMP) in a graph. We propose a heuristic algorithm to solve the OMP problem in this new context and compare its results with those of the exact solution algorithm, showing its low error rate and short runtime. Finally, we propose some exploratory directions for future research.

Neural networks have a proven usefulness at predicting, denoising or classifying time series. However, the performance of deep learning models is bound to the size of the input window. Yet, no common method has emerged to determine the optimal window size. In this paper, we compare two heuristics and three event detection algorithms to find the best time representation for three different tasks, using one simulated and two real-world datasets. The two real-world applications are the electricity disaggregation for energy efficiency in buildings and the detection of fibrillation for diagnosis in cardiology. We compare the obtained window sizes with the experimental values from previous research and we experimentally validate the relevance of the results using both convolutional and recurrent deep neural networks. Results confirm the impact of the sequence length on model performance and show that window sizes cannot be simply transferred to another dataset, even for the same problem. We also find that the false nearest neighbors method can reliably estimate the window size and can help with the tedious work of finding the right time representation.

This paper presents a novel and unconventional NAS method called DensEMANN, which can automatically build DenseNet and DenseNet-BC architectures layer by layer and kernel by kernel while they are being trained. DensEMANN is a complex method, divided into a macro-algorithm (which adds dense layers one by one to the network) and a micro-algorithm (which builds new layers by adding and/or pruning convolution kernels from them). Using the micro-algorithm alone, it is possible to build one-layer DenseNet trained on CIFAR-10 and SVHN with similar or greater accuracy levels than those obtained with equivalent prebuilt architectures. As for the macro-algorithm, its results are still unpredictable, and further research is needed before its automatically-generated DenseNet can compete with neural networks designed by humans.

This paper presents a newly developed self-structuring algorithm for generating convolutional neural networks, as well as the results of preliminary tests performed on it. The algorithm produces DenseNet and DenseNet-BC architectures layer by layer from scratch, at the same time as they are being trained. Experimental results for well-known image classification datasets (CIFAR-10 and SVHN) are promising. The accuracy levels of generated networks are not significantly different than those of prebuilt DenseNet and DenseNet-BC with similar topologies, and are approaching the state of the art for these datasets.

Atrial fibrillation (AF) is the most common heart arrhythmia. It affects between 1% and 2% of the world population over 35 years old. This disease is linked to an increased risk of stroke and heart failure. AF is a progressive disease and, at first, paroxysmal AF episodes occur, last from seconds up to a week and then stop. The disease evolves to permanent state, where the heart is always in fibrillation and can't be corrected. Forecasting paroxysmal AF episode a few seconds or minutes before its onset remains a hard challenge, but could lead to new treatment methods. For this study, we constructed a new long-term electrocardiogram (ECG) database (24 to 96 hours), composed of 10484 ECG. As a result of a careful analysis by a cardiologist, 250 AF onsets of paroxysmal AF have been detected in 140 ECG. We developed a deep neural network (DNN) model, composed of convolutional neural network (CNN) layers and bidirectional gated recurrent units (GRU) as recurrent neural network (RNN) layers. The model was trained for a supervised binary classification distinguishing between heartbeats series (RR intervals) that precede an AF onset and series distant from any AF. The model achieved an average area under the receiver operating characteristic (ROC) curve of 0.74. We evaluated the impact of heartbeat window size given as input, and the time period between the heartbeats window and the AF onset. We found that an input window of 300 heartbeats gives the best results and, not surprisingly, the closer the window is from the AF onset, the better the results. We concluded that RR intervals series contains information about the incoming AF episode, and that it can be exploited to forecast such episode.

We present a clustering method designed to complement existing model-based collaborative filtering algorithms. Thanks to clustering, we are able to produce faster recommendations on datasets with a very large number of items. Moreover, we can work with models independent of the number of users, and produce recommendations for new users and in session based environment. We show that our method builds meaningful clusters, and that we can reduce the number of items considered during recommendation by a factor up to 50, with little impact on the accuracy or diversity of the recommendation.

Recurrent neural networks have recently been successfully applied to the session-based recommendation problem, and is part of a growing interest for collaborative filtering based on sequence prediction. This new approach to recommendations reveals an aspect that was previously overlooked: the difference between short-term and long-term recommendations. In this work we characterize the full short-term/long-term profile of many collaborative filtering methods, and we show how recurrent neural networks can be steered towards better short or long-term predictions. We also show that RNNs are not only adapted to session-based collaborative filtering, but are perfectly suited for collaborative filtering on dense datasets where it outperforms traditional item recommendation algorithms.

In the history of agent-based/Alife simulations helping to better understand some unexpected social consequences emerging out of interactive individual behaviors, two models have become quite celebrated: the Schelling’s segregation model and the Maynard-Smith’s emergence of cooperation in an albeit very defective world. In this paper, we explore a simulation at the crossroad of these two models. We show how in a bi-color society in which individual can move according to their neighborhood “colored” composition (reminiscent of the original Schelling’s model), one further way to favor cooperation is to encourage “communitarian” behavior i.e. allowing individual to move, compose clusters of similar color and evolve a cooperative behavior restricted to partners of the same color. In brief, our results not so surprisingly tend to show that communitarian cooperation and segregation evolve hand in hand.

Although criticized for some of its limitations, modularity remains a standard measure for analyzing social networks. Quantifying the statistical surprise in the arrangement of the edges of the network has led to simple and powerful algorithms. However, relying solely on the distribution of edges instead of more complex structures such as paths limits the extent of modularity. Indeed, recent studies have shown restrictions of optimizing modularity, for instance its resolution limit. We introduce here a novel, formal and well-defined modularity measure based on random walks. We show how this modularity can be computed from paths induced by the graph instead of the traditionally used edges. We argue that by computing modularity on paths instead of edges, more informative features can be extracted from the network. We verify this hypothesis on a semi-supervised classification procedure of the nodes in the network, where we show that, under the same settings, the features of the random walk modularity help to classify better than the features of the usual modularity. Additionally, the proposed approach outperforms the classical label propagation procedure on two data sets of labeled social networks.

This special issue pays tribute to 20 years of artificial life research in Europe. Since its inception around a coffee table in Paris in the summer of 1990, the European Conference on Artificial Life (ECAL) has grown and evolved to incorporate multiple research disciplines that aim to combine the natural and computational sciences. In addition to extending a long and rich tradition in theoretical biology present at that time in Europe toward the computational sciences and engineering, ECAL has also provided a platform for different innovative ideas that have produced a significant impact on other research domains or have resulted in novel research directions with their own workshops and conferences. This influence continues today, with the consideration of the different models and methods that are currently extensively used to answer biological questions in domains such as computational and systems biology. This evolution will benefit the field and will ensure that our models continue to push the envelope as they have been doing for the last two decades.Together with the reviewers and conference chairs of ECAL 2011 in Paris, we selected a small number of especially interesting contributions among the conference's 128 accepted submissions. Their authors were asked to extend their manuscripts for a wider audience and to provide more details than what was allowed by the conference format. The reader can find below a brief description of these articles, which range from artificial chemistry to evolutionary robotics and computational biology. We hope that this selection will provide an incentive for researchers from related fields to pay closer attention to the innovations created within the artificial life community and to motivate researchers already interested in this field to make further advances in the realism and influence of their models. In the following we summarize the topics discussed in this special issue.Xabier Barandiaran and Matthew Egbert study how norms become established from an organism-centered perspective. They give an in-depth analysis and evaluation of a minimal model capable of establishing norms dynamically. The model clearly opens a novel road toward answering questions associated to the notion of minimal agency, which is essential to a wide variety of artificial life models.The work of Navneet Bhalla, Peter Bentley, Peter Vize, and Christian Jacob focuses on the use of staging in the context of physical self-assembling systems made of magnetic shapes. They investigate whether replacing a full mix with gradual addition of morphological information manages to produce more stable constructs for systems composed of elements with heterogeneous properties. They show that this staging methodology can solve complex self-assembly problems by reducing the possibilities of errors in the shape-matching process.Tom Froese, Nathaniel Virgo, and Takashi Ikegami argue in the third article that time scales should play a central role in all models aiming to answer questions linked to the origin of life. Using a reaction-diffusion model, they assert that adaptive behavior realized at intermediate time scales by spatiotemporally localized self-producing systems may have been already present in the beginnings, and conclude with the feasibility of a movement-first approach to the origin of life.In the fourth contribution, Heiko Hamann, Thomas Schmickl, and Karl Crailsheim investigate the relevance of statistical mechanics, and more specifically of the fluctuation theorem, to understanding the collective behavior of simulated swarms. They are able to show that an inverted fluctuation theorem is present in this type of systems, which they claim may be generally applicable to different self-organizing systems studied in artificial life.Tsutomu Oohashi, Tadao Maekawa, Osamu Ueno, Norie Kawai, Emi Nishina, and Manabu Honda investigate a completely different question: Can organisms evolve to give up their immortality? Using an artificial-chemistry-based model called SIVA, they examine the evolution of programmed self-decomposition. Interestingly, they identify the conditions for mortal organisms emerging from immortal ones to outcompete them, which they associate with the emergence of an altruistic activity in terrestrial life.As mentioned earlier, artificial life modeling has a prominent role to play in computational and systems biology. The manuscript of Joshua Payne, Jason Moore, and Andreas Wagner is an example of this cross-fertilization. The authors investigate the space of signal-integration functions, which are defined as functions that map regulatory signals to gene expression states. Their aim is to understand the relationship between the robustness and evolvability of gene expression states when genetic perturbations modify these functions. They show, among other things, that the majority of signal-integration functions are resistant to perturbation.José Pereira, Porfírio Silva, Pedro Lima, and Alcherio Martinoli propose in their work a new framework to coordinate the actions of robot teams using concepts from the field of institutional economics. They first introduce the notion of executable Petri nets (EPNs), which can be run directly on a robot. These EPNs are then used to define institutions and individual robot behaviors, resulting in what they call an institutional agent controller. To validate this new approach, they design some swarm-related experiments in which robots need to coordinate to remain connected over a wireless network, showing that their novel institutional robotics framework can perform as well as some of the other approaches that aim to design swarm behavior, while offering a higher potential for modular control design and consideration of social behavior in robot teams.In the eighth contribution to this special issue, John Rieffel, Davis Knox, Schuyler Smith, and Barry Trimmer provide insight into four years of research into soft robotics. They particularly focus on the question of how to coevolve soft robot morphology and control. They argue that the evolution of soft robots corresponds to solving a problem of three tightly interdependent variables, namely material, morphology and control. In their study, they provide three approaches to address this problem, showing a way to achieve physically embodied soft robots.Within the 20 years of the European Conference of Artificial Life, numerous articles have addressed new evolutionary algorithms designed to solve particular problems or to overcome specific issues in the model originally proposed by John Holland. The ninth and last article, by Nicholas Tomko, Inman Harvey, Nathaniel Virgo, and Andrew Philippides, focuses here on the question of how to improve on prior work about niching and speciation. They suggest a new genetic algorithm (GA), called the group GA, which is based on principles of group or multilevel selection. They show that this new algorithm outperforms some of the earlier works in the same context, and that the group size need not be preset but can itself be evolved.

Genomic data integration is a key goal to be achieved towards large-scale genomic data analysis. This process is very challenging due to the diverse sources of information resulting from genomics experiments. In this work, we review methods designed to combine genomic data recorded from microarray gene expression (MAGE) experiments. It has been acknowledged that the main source of variation between different MAGE datasets is due to the so-called 'batch effects'. The methods reviewed here perform data integration by removing (or more precisely attempting to remove) the unwanted variation associated with batch effects. They are presented in a unified framework together with a wide range of evaluation tools, which are mandatory in assessing the efficiency and the quality of the data integration process. We provide a systematic description of the MAGE data integration methodology together with some basic recommendation to help the users in choosing the appropriate tools to integrate MAGE data for large-scale analysis; and also how to evaluate them from different perspectives in order to quantify their efficiency. All genomic data used in this study for illustration purposes were retrieved from InSilicoDB http://insilico.ulb.ac.be.

The potential of microarray gene expression (MAGE) data is only partially explored due to the limited number of samples in individual studies. This limitation can be surmounted by merging or integrating data sets originating from independent MAGE experiments, which are designed to study the same biological problem. However, this process is hindered by batch effects that are study-dependent and result in random data distortion; therefore numerical transformations are needed to render the integration of different data sets accurate and meaningful. Our contribution in this paper is two-fold. First we propose GENESHIFT, a new nonparametric batch effect removal method based on two key elements from statistics: empirical density estimation and the inner product as a distance measure between two probability density functions; second we introduce a new validation index of batch effect removal methods based on the observation that samples from two independent studies drawn from a same population should exhibit similar probability density functions. We evaluated and compared the GENESHIFT method with four other state-of-the-art methods for batch effect removal: Batch-mean centering, empirical Bayes or COMBAT, distance-weighted discrimination, and cross-platform normalization. Several validation indices providing complementary information about the efficiency of batch effect removal methods have been employed in our validation framework. The results show that none of the methods clearly outperforms the others. More than that, most of the methods used for comparison perform very well with respect to some validation indices while performing very poor with respect to others. GENESHIFT exhibits robust performances and its average rank is the highest among the average ranks of all methods used for comparison.

BACKGROUND: With an abundant amount of microarray gene expression data sets available through public repositories, new possibilities lie in combining multiple existing data sets. In this new context, analysis itself is no longer the problem, but retrieving and consistently integrating all this data before delivering it to the wide variety of existing analysis tools becomes the new bottleneck. RESULTS: We present the newly released inSilicoMerging R/Bioconductor package which, together with the earlier released inSilicoDb R/Bioconductor package, allows consistent retrieval, integration and analysis of publicly available microarray gene expression data sets. Inside the inSilicoMerging package a set of five visual and six quantitative validation measures are available as well. CONCLUSIONS: By providing (i) access to uniformly curated and preprocessed data, (ii) a collection of techniques to remove the batch effects between data sets from different sources, and (iii) several validation tools enabling the inspection of the integration process, these packages enable researchers to fully explore the potential of combining gene expression data for downstream analysis. The power of using both packages is demonstrated by programmatically retrieving and integrating gene expression studies from the InSilico DB repository [https://insilicodb.org/app/].

A plenitude of feature selection (FS) methods is available in the literature, most of them rising as a need to analyze data of very high dimension, usually hundreds or thousands of variables. Such data sets are now available in various application areas like combinatorial chemistry, text mining, multivariate imaging, or bioinformatics. As a general accepted rule, these methods are grouped in filters, wrappers, and embedded methods. More recently, a new group of methods has been added in the general framework of FS: ensemble techniques. The focus in this survey is on filter feature selection methods for informative feature discovery in gene expression microarray (GEM) analysis, which is also known as differentially expressed genes (DEGs) discovery, gene prioritization, or biomarker discovery. We present them in a unified framework, using standardized notations in order to reveal their technical details and to highlight their common characteristics as well as their particularities.

Autocatalysis is a fundamental concept, used in a wide range of domains. From its most general definition, that is, a process in which a chemical compound is able to catalyze its own formation, several different systems can be described. We detail the different categories of autocatalyses, and compare them on the basis of their mechanistic, kinetic, and dynamic properties. It is shown how autocatalytic patterns can be generated by different systems of chemical reactions. The notion of autocatalysis covers a large variety of mechanistic realizations with very similar behaviors; it is proposed that its key signature is its kinetic pattern expressed in a mathematical form. This notion, while describing dynamic behaviors at the most fundamental level, is at the basis for developing higher-level concepts towards life: autocatalytic sets, and autopoietic systems.

Abstract Microarray technology has become an integral part of biomedical research and increasing amounts of datasets become available through public repositories. However, re-use of these datasets is severely hindered by unstructured, missing or incorrect biological samples information; as well as the wide variety of preprocessing methods in use. The inSilicoDb R/Bioconductor package is a command-line front-end to the InSilico DB, a web-based database currently containing 86 104 expert-curated human Affymetrix expression profiles compiled from 1937 GEO repository series. The use of this package builds on the Bioconductor project's focus on reproducibility by enabling a clear workflow in which not only analysis, but also the retrieval of verified data is supported. Availability: inSilicoDb is available as part of the Bioconductor project. There is a companion web interface that can be used for browsing available datasets before importing them into R/Bioconductor (http://insilico.ulb.ac.be). Contact: [email protected]

While fixed point dynamics is still the predominant regime used for information processing, recent brain observations and computational results suggest more and more the importance/necessity to include and rely on more complex dynamics. Independently, since their introduction sixty years ago, cell assemblies are still a powerful substrate for brain information processing. Here, the first part of this paper aims to conciliate these two evidences by investigating the possibility to encode content addressable information in pre-encoded cell assemblies characterized by complex dynamics. As an expected outcome, after stimulus offset, the information is maintained in the attractor of the cell assembly. As a less expected outcome, when the system is fed with ambiguous stimuli, it will continuously iterate across the possible attractors (instead to settle down to a specific one). In the second part of the paper, based on biologically plausible mechanisms, a novel unsupervised algorithm for online cell assemblies creation is proposed. The procedure involves simultaneously, a fast hebbian/anti-hebbian learning of the network's recurrent connections for the creation of new cell assemblies, and a slower feedback signal which stabilizes the cell assemblies by learning the feedforward input connections. Results show that the obtained cell assemblies exhibit similar behavior as the pre-encoded ones. Finally, we propose that this model could be working for the cognitive map formation of multiple place fields in the CA3 network when the rat is facing a new environment.

A coevolutionary model is discussed that incorporates the logical structure of constitutional chemistry and its kinetics on the one hand and the topological evolution of the chemical reaction network on the other hand. The motivation for designing this model is twofold. First, experiments that are to provide insight into chemical problems should be expressed in a syntax that remains as close as possible to real chemistry. Second, the study of physical properties of the complex chemical reaction networks requires growing models that incorporate features realistic from a biochemical perspective. In this article the theory and algorithms underlying the coevolutionary model are explained, and two illustrative examples are provided. These examples show that one needs to be careful in making general claims concerning the structure of chemical reaction networks.

The basic idea to defend in this paper is that an adequate perception of the search space, sacrificing most of the precision, can paradoxically accelerate the discovery of the most promising solution zones. While any search space can be observed at any scale according to the level of details, there is nothing inherent to the classical metaheuristics to naturally account for this multi-scaling. Nevertheless, the wider the search space the longer the time needed by any metaheuristic to discover and exploit the "promising" zones. Any possibility to compress this time is welcome. Abstracting the search space during the search is one such possibility. For instance, a common Ordering Genetic Algorithm (o-GA) is not well suited to efficiently resolve very large instances of the Traveling Salesman Problem (TSP). The mechanism presented here (reminiscent of Gestalt psychology) aims at abstracting the search space by substituting the variables of the problems with macro-versions of them. This substitution allows any given metaheuristic to tackle the problem at various scales or through different multi-resolution lenses. In the TSP problem to be treated here, the towns will simply be aggregated into regions and the metaheuristics will apply on this new one-level-up search space. The whole problem becomes now how to discover the most appropriate regions and to merge this discovery with the running of the o-GA at the new level.

How information and more particularly memories are represented in brain dynamics is still an open question. By using, a recurrent network receiving a stimulus dependent external input, the author have demonstrated that the use of limit cycle attractors encompass in many aspects the limitations of fixed points attractors and gives better correspondence with neurophysiological facts. A main outcome of this perspective is the apparition of chaotic trajectories: instead of the overwhelming presence of spurious attractors, chaotic dynamics shows up when facing ambiguous situation. Contrary to intuition, many studies reported that noise can have beneficial effects in dynamical systems. Inline with these studies, it is demonstrated here how stochastic noise can make converge the chaotic trajectories to the expected limit cycle attractors and accordingly can improve consequently the retrieval performance. This noise induced retrieval enhancement is very dependent of the type of chaotic dynamics which is function of how information is coded.

This letter aims at studying the impact of iterative Hebbian learning algorithms on the recurrent neural network's underlying dynamics. First, an iterative supervised learning algorithm is discussed. An essential improvement of this algorithm consists of indexing the attractor information items by means of external stimuli rather than by using only initial conditions, as Hopfield originally proposed. Modifying the stimuli mainly results in a change of the entire internal dynamics, leading to an enlargement of the set of attractors and potential memory bags. The impact of the learning on the network's dynamics is the following: the more information to be stored as limit cycle attractors of the neural network, the more chaos prevails as the background dynamical regime of the network. In fact, the background chaos spreads widely and adopts a very unstructured shape similar to white noise. Next, we introduce a new form of supervised learning that is more plausible from a biological point of view: the network has to learn to react to an external stimulus by cycling through a sequence that is no longer specified a priori. Based on its spontaneous dynamics, the network decides "on its own" the dynamical patterns to be associated with the stimuli. Compared with classical supervised learning, huge enhancements in storing capacity and computational cost have been observed. Moreover, this new form of supervised learning, by being more "respectful" of the network intrinsic dynamics, maintains much more structure in the obtained chaos. It is still possible to observe the traces of the learned attractors in the chaotic regime. This complex but still very informative regime is referred to as "frustrated chaos."

Most metaheuristics try to find a good balance between exploitation and exploration to achieve their goals. The exploration efficiency is highly dependent on the size and ruggedness of the search space. A metaheuristic such as the Simple Genetic Algorithm (SGA) is not totally suited to traverse very large landscapes, especially deceptive ones. The approach proposed here improves the exploration of the SGA by adding a second search process through the way solutions are coded. We will designate by “ observer” each possible coding that aims at reducing the search space. Information on the quality and adequacy of one possible observer is obtained by adopting this specific coding and testing how SGA benefits from this search space reduction. The new approach investigated here trains the observers for a specific time and then decides which of them is the most suitable to solve the whole problem. Concretely, a second evolutionary stage has been added to evolve observers for the SGA. These observers aim at reducing the size and smoothing the ruggedness of the search space through a simplification of the genotype. To test the proposed approach, we apply it to the classical Hierarchical IF- and-only-iF (HIFF) problem and measure the efficiency of our approach in term of quality and time.

Many metaheuristics have difficulty exploring their search space comprehensively. Exploration time and efficiency are highly dependent on the size and the ruggedness of the search space. For instance, the Simple Genetic Algorithm (SGA) is not totally suited to traverse very large landscapes, especially deceptive ones. The approach introduced here aims at improving the exploration process of the SGA by adding a second search process through the way the solutions are coded. An "observer" is defined as each possible encoding that aims at reducing the search space. Adequacy of one observer is computed by applying this specific encoding and evaluating how this observer is beneficial for the SGA run. The observers are trained for a specific time by a second evolutionary stage. During the evolution of the observers, the most suitable observer helps the SGA to find a solution to the tackled problem faster. These observers aim at collapsing the search space and smoothing its ruggedness through a simplification of the genotype. A first implementation of this general approach is proposed, tested on the Shuffled Hierarchical IF-and-only-iF (SHIFF) problem. Very good results are obtained and some explanations are provided about why our approach tackles SHIFF so easily.

Seminal observations performed by Skarda and Freeman (1987) on the olfactory bulb of rabbits during cognitive tasks have suggested to locate the basal state of behavior in the network's spatio-temporal dynamics. Following these neurophysiological observations, the authors have investigated in previous papers the possibility to store external stimuli in spatio-temporal dynamical attractors of recurrent neural networks. To this aim, an efficient learning algorithm, based on a time asymmetric Hebbian mechanism, has been proposed. The underlying idea is to obtain - as much as possible - a natural i.e. unconstrained mapping between the external stimuli and the spontaneous internal dynamics of the network. The dynamical regime called "frustrated chaos" by the authors appears to play a substantial role in the establishment of this mapping. In this paper, adopting a symbolic coding of the output, new investigations are performed on the presence and the importance of spurious data. It is shown how the presence of chaos contributes to stop their proliferation.

No abstract available.

This paper aims at studying the impact of an Hebbian learning algorithm on the recurrent neural network's underlying dynamics. Two different kinds of learning are compared in order to encode information in the attractors of the Hopfield neural net: the storing of static patterns and the storing of cyclic patterns. We show that if the storing of static patterns leads to a reduction of the potential dynamics following the learning phase, the learning of cyclic patterns tends to increase the dimension of the potential attractors instead. In fact, such learning may be used as an extra "route to chaos": the more cycles to be learned, the more the network shows as spontaneous dynamics a form of chaotic itinerancy among brief oscillatory periods. These results are in line with the observations made by Freeman in the olfactory bulb of the rabbit: cycles are used to store information and the chaotic dynamics appears as the background regime composed of those cyclic "memory bags". It confirms precedent papers in which it was observed that huge encoding capacity in term of cyclic attractors implies strong presence of chaos.

The learning impact, of an iterative supervised Hebbian learning algorithm, on a recurrent neural network's underlying dynamics has been discussed in a previous paper. It was argued that these results are in line with the observations made by Freeman in the olfactory bulb of the rabbit: cycles are used to store information and the chaotic dynamics appears as the background regime composed of those cyclic "memory bags". However, to get closer to a biological point of view, this paper introduces an unsupervised version of this Hebbian algorithm. As a direct result, both the storing capacity and the content addressability of the learned networks are greatly enhanced. Furthermore, stunning dynamical results are observed: if the learning process increases the dimension of the potential attractors, however, less chaoticity is found than in a supervised learning process. Moreover, chaos obtained looks more structured, made from brief itinerancy among learned cycles.

So far, recurrent networks, when adopting fixed point dynamics, show a very poor encoding capacity. However, these same networks, when preferentially maintained in chaotic dynamics, can encode an enormous amount of information in their cyclic attractors and this boosts their encoding capacity. It has been described in a previous paper a simple way to encode such information by robustly associating each vector in a N-dimensional space with one "symbolic" cyclic attractor. The main message was the monotonous increase of chaotic spontaneous regimes as a function of the number of attractors to learn. However, no algorithm was provided to adjust the connection's weight in order to encode a given input set. For this purpose, this paper revisits the classical gradient-based BPTT learning algorithm. It shows that this algorithm gives poor results and furthermore that by using it the "chaoticity" of the network dampens strongly, hence it's encoding capacity.

This papers aims at experimentally confirming in very small Hopfield networks that chaos becomes the spontaneous dynamics of a network, when the number of inputs to be learned and stored in its dynamical attractors is superior to the size of this network. The spontaneous dynamics of the network are shown to increase in complexity by increasing the size of the learning set. The type of chaos exploited to code these inputs is related with the frustrated chaos described in previous papers. When, following a brute-force but robust learning of the inputs, the network is presented with a series of unlearned ambiguous input, experimental results show how chaotic regimes become the natural response of the network, reflecting in its frustrated dynamics the inherent ambiguity of the input. A live demonstration can be shown where the rhythms associated with the dynamics can be heard. This experimental work might give additional support to the Skarda and Freeman strong intuition that chaos should play an important role in the storage and the search capacities of our brains.

This paper carries out the study of natural networks (genetic, chemical and immune) evolving according to two levels of change called dynamics and metadynamics. The dynamics is the evolution in time of the concentration of the units currently present in the networks: the genetic, the molecular or the immune species. Their concentration evolves as a function of their network interaction with the other units. However, this evolution is also a function of their "exogenous" fitness so that the fitter units should in principle grow faster than the others. The metadynamics is the only way for innovation, and amounts to the generation of new units on the basis of the genetic or chemical materials constituting the units existing so far in the network. This metadynamics, indirectly subject to the exogenous pressure, tends to selectively favor units that are easier to produce from the existing ones. For instance, the genetic recombination of two species could occur between species presenting particular similar properties and thus generating new species merging these properties. This metadynamics also greatly influences the concentration of the units present in the network. The paper experimentally shows, on chemical, genetic and immune networks, that the interaction between these two levels of change, together with the intricate balance between the "exogenous" and the "network endogenous" selective drift, can induce a hard-to-predict concentration profile, subject to discontinuous changes.

SCOPUS: ar.j

Fuzzy controllers are now part of the common control tools, and several different approaches for their design have been proposed in the literature. This paper presents a direct adaptive fuzzy controller for unknown nonlinear systems. Starting from the single step ahead direct adaptive control approach, a multiple steps ahead direct adaptive control scheme is presented. These approaches have been both developed at IRIDIA in the framework of the FAMIMO Esprit project. The strength of the second method is to overcome the problems related with the first approach, such as its inability to control nonminimum phase systems, at the price of increased computational load, and a deeper knowledge of the model of the system.

Backpropagation-through-time (BPTT) is the temporal extension of backpropagation which allows a multilayer neural network to approximate an optimal state-feedback control law provided some prior knowledge (Jacobian matrices) of the process is available. In this paper, a simplified version of the BPTT algorithm is proposed which more closely respects the principle of optimality of dynamic programming. Besides being simpler, the new algorithm is less time-consuming and allows in some cases the discovery of better control laws. A formal justification of this simplification is attempted by mixing the Lagrangian calculus underlying BPTT with Bellman-Hamilton-Jacobi equations. The improvements due to this simplification are illustrated by two optimal control problems: the rendezvous and the bioreactor.

Hopfield network transient dynamics have been exploited for resolving both path planning and temporal pattern classification. For these problems Lagrangian techniques and two well-known learning algorithms for recurrent networks have been used. For path planning, the Williams and Zisper's learning algorithm has been implemented and a set of temporal trajectories which join two points, pass through others, avoid obstacles and jointly form the shortest path possible are discovered and encoded in the weights of the net. The temporal pattern classification is based on an extension of the Pearlmutter's algorithm for the generation of temporal patterns which is obtained by means of variational methods. The algorithm is applied to a simple problem of recognizing five temporal trajectories with satisfactory robustness to distortions.