Artificial Neural Networks have been crowned with tremendous successes in recent years, with ever wider and more complex ranges of applications. However, they, too often, result from a costly human design process relying as much on expertise as on trial and error. While the field of NeuroEvolution provides a complementary view point through emergent, self-designing ANNs, the “black-box” properties of the resulting networks is further magnified. Still, by once more taking inspiration from biology, we may extract meaningful information from ANNs by using similar approaches as those used for biological brains. In this work, we study the emergence and functional allocation of neurons in a light communication task. By having a robot transmit visual information, through vocal channels, we enrich the existing literature with new types of stimuli, namely those related to role (emitter/ receiver). Through Virtual functional Magnetic Resonance Imaging (VfMRI), we observe that evolution only favored specific kind of input-processing modules. Combined with a strong presence of jack-of-alltrades modules, this demonstrates the balancing act between specialization and generalization in Artificial Neural Networks with emergent topologies.

Understanding the neurological implementation of emotions is a major research subject from biology to computer sciences that, in the latter case, takes many shapes: from accurate detection of human emotions to the emulation of plausible responses to stimuli. There is, however, room for a more bottom-up approach in which we would thrive to recreate emotions from undifferentiated elementary building blocks. In this article, we used virtual creatures that interact with their environment through a low-level perception/cognition/action loop to demonstrate their potential for fear responses. Embedded in a physical environment in a typical prey/predator setting, they develop strategies for foraging while minimizing their exposure to danger. By monitoring the neural activities of these subjects, we were able to highlight the regularities induced by an ES-HyperNEAT encoding and their eventual mapping into “mental states”. We further emphasize the potential of this approach by clustering these ANNs and showing their resulting complexity in terms of conspecific identification, communication, and functional modularity. Indeed, through functional equivalence across numerous topologies, we identify a fear-related neural cluster that serves as a primitive defensive survival circuit.

Progress in molecular genetics allowed taxonomists to better understand the relationships between species without the bias of morphological similarities. However, access to data from times past is limited to the fossil archives which, being far from complete, can only provide limited information. To address this problem through the field of Artificial Life, we devised a polyvalent sexual reproduction scheme and an automated phylogenetic tool capable of producing, from a stream of genomes, hierarchical species trees with relatively low memory footprint. We assert that these apparatus perform well under reasonable stress by embedding them into 2D simulations of unsupervised plant evolution in textbook cases of geographical speciation. After thousands of generations and millions of plants, the extracted phylogenetic data not only showed the expected results in terms of branching pattern (anagenesis, cladogenesis) but also exhibited complex interactions between species both in space and time.

One of the challenges of researching spiking neural networks (SNN) is translation from temporal spiking behavior to classic controller output. While many encoding schemes exist to facilitate this translation, there are few benchmarks for neural networks that inherently utilize a temporal controller. In this work, we consider the common reinforcement problem of animat locomotion in an environment suited for evaluating SNNs. Using this problem, we explore novel methods of reward distribution as they impacts learning. Hebbian learning, in the form of spike time dependent plasticity (STDP), is modulated by a dopamine signal and affected by reward-induced neural activity. Different reward strategies are parameterized and the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) is used to find the best strategies for fixed animat morphologies. The contribution of this work is two-fold: to cast the problem of animat locomotion in a form directly applicable to simple temporal controllers, and to demonstrate novel methods for reward modulated Hebbian learning.

We present the use of a new computationaly efficient 3D physics model for the simulation of cells in a virtual aquatic world. In this model, cells can freely assemble and disconnect along the simulation without any separation between the development and evaluation stages, as is the case in most evo-devo models which only consider one cell cluster. While allowing for the discovery of interesting behaviors through the addition of new degrees of freedom, this 3D center-based physics engine and its associated virtual world also come with their drawbacks when applied to evolutionnary experiments: larger search space and numerous local optima. In this paper, we have designed an experiment in which cells must learn to survive by keeping their genome alive as long as possible in a demanding world. No morphology or strategy is explicitly enforced; the only objective the cells have to optimize is the survival time of the organism they build. We show that a novelty metric, adapted to our evo-devo matter, dramatically improves the outcome of the evolutionary runs. This paper also details some of the developmental strategies the evolved multicellular organisms have found in order to survive.