Catálogo de publicaciones - libros

It is speculated that there is a relationship between 1/ noise and computational universality in cellular automata. We use genetic algorithms to search for one-dimensional and two-state, five-neighbor cellular automata which have 1/-type spectrum. A power spectrum is calculated from the evolution starting from a random initial configuration. The fitness is estimated from the power spectrum in consideration of the similarity to 1/-type spectrum. The result shows that the rule with the highest average fitness has a propagating structure like other computationally universal cellular automata, although computational universality of the rule has not been proved yet.

The relative influence of genotype reuse and number of genotype parameters on the evolvability of an embodied neural network is explored. Two genotype to phenotype mappings are used to encode a neural network controlling a hexapod agent. A symmetric encoding reuses the genotype by duplicating parts of the genotype to create the phenotype. A direct encoding maps one genotype parameter to one phenotype parameter. To test whether genotype reuse is more important than genotype size, the architecture of the neural network is manipulated such that the genotype size of the symmetrically-encoded neural networks is larger than the directly-encoded neural networks. The symmetrically-encoded neural networks are found to be more evolvable than the directly-encoded despite having a larger genotype.

In this paper, we introduce a method, Assortative Preferential Attachment, to grow a scale-free network with a given assortativeness value. Utilizing this method, we investigate information-cloning — recovery of scale-free networks in terms of their information transfer — and identify a number of recovery features: a full-recovery threshold, a phase transition for both assortative and disassortative networks, and a bell-shaped complexity curve for non-assortative networks. These features are interpreted with respect to two opposing tendencies dominating network recovery: an increasing amount of choice in adding assortative/disassortative connections, and an increasing divergence between the joint remaining-degree distributions of existing and required networks.

A novel approach to topology and weight evolving artificial neural networks (TWEANNs) is presented. Compared with previous TWEANNs, this method has two major characteristics. First, a set of genetic operations may be designed without recombination because it often generates an offspring whose fitness value is considerably worse than its parents. Instead, two topological mutations whose effect on fitness value is assumed to be nearly neutral are provided in the genetic operations set. Second, a new encoding technique is introduced to define a string as a set of substrings called operons. To examine our approach, computer simulations were conducted using the standard reinforcement learning problem known as the double pole balancing without velocity information. The results obtained were compared with NEAT results, which is recognised as one of the most powerful techniques in TWEANNs. It was found that our proposed approach yields competitive results, especially when the problem is difficult.

We present a model of a recurrent neural network with homeostasic units, embodied in a minimalist articulated agent with a single link and joint. The configuration of the agent is determined by the total activation level or kinetic energy of the network. We study the complexity patterns of the neural networks, and see how the entropy of the neural controller state and agent configuration changes with the relative characteristic time of the homeostasis when compared with the excitatory-inhibitory activation dynamics of network. We also present a meta-model of embodied neural agents, that serves as conceptual framework to study self-perturbation and the self-organization in embodied neural agents. Simulation results show that homeostasis significantly influences the dynamics of the network and the controlled agent, allowing the system to escape fixed-points and produce complex aperiodic behavior. The relation between the characteristic time of homeostasis and the characteristic time of main excitatory-inhibitory activation dynamics was found to be non-linear and non-monotonic. We use these findings to connect the perspectives of classical cybernetics on homeostasis to complexity research.

We propose a generalized framework to analyse constraints and representations in growing complex networks. We show that the introduction of biological, social and technological information by means of an additional network of constraints, together with the distinction between complete potential networks and instantaneous effective ones, can offer additional insights about the final topological outcome. Specifically, we study the emergence of exponential cutoffs in broad-scale degree distributions as a result of high level constraints.

This paper addresses the role of space in evolving a novel Non-Spatial GasNet model. It illustrates that this particular neural network model which make use of modulatory effects of diffusing gases has its evolvability improved when its neurons are not constrained to a Euclidean space. The results show that successful behaviour is achieved in fewer evaluations for the novel unconstrained GasNet than for the original model.

We study the dynamics of Boolean networks of the scale-free type. The model takes into account the topology and abstracts recent findings about real genetic regulatory networks. We propose a new, more biologically plausible, semi-synchronous update scheme on networks of larger sizes. We simulate statistical ensembles of networks and discuss the attractors of the dynamics, showing that it is compatible with theoretical biological network models. Moreover, then model demonstrates interesting scaling abilities as the size of the networks is increased.

The “order for free” exhibited by some classes of system has been exploited by natural selection in order to build systems capable of exhibiting complex behaviour. Here we explore the impact of one ordering constraint, spatial embedding, on the dynamical complexity of networks. We apply a measure of functional complexity derived from information theory to a set of spatially embedded network models in order to make some preliminary characterisations of the contribution of space to the dynamics (rather than mere structure) of complex systems. Although our measure of dynamical complexity hinges on a balance between functional integration and segregation, which seem related to an understanding of the small-world property, we demonstrate that small-world structures alone are not enough to induce complexity. However, purely spatial constraints can produce systems of high intrinsic complexity by introducing multiple scales of organisation within a network.

The article suggests new ideas regarding computer simulations in pedestrian dynamics, which take into account interaction between particular pedestrians. The described situation is a room evacuation for chosen classes of situations: normal room evacuation, controlled evacuation, and panic. Based on the formalism of Extended Cellular Automata, the following models of pedestrian dynamics are presented: a basic model, SPA (Strategic Pedestrian Abilities) model, SPA–BNE (Bottleneck Effect) model, which were created with reference to the particular classes of situations.