Abstract:  When using artificial evolution to design AI systems, you are forced to make very explicit, in operational and testable terms, what properties you wish to select. If these properties relate to interactions with humans, then you must weigh up the costs of evolving in the real world against the perils of using simulations. The financial crash is a salutary reminder of the difference between models and the real world, and how common intuitions can mislead when applied to complex systems.

Is trust in tools and in people a social phenomenon? What are the key features of the context in which SmartData avatars would need to operate? Is there a minimal toy scenario to demonstrate the key issues? Is there an evolutionary pathway from such toy worlds to the messy complex real world? This talk will present more questions than answers.

Abstract:  Understanding the evolution of communication and human language is one of the hardest problems in science. Of significant research and practical interest is the related artificial perspective: How can populations of artificial agents evolve/develop forms of communication of varying complexity, analogous to animal and human communication? Recent progress in this area permit us to progress our understanding of long-debated issues such as how symbols and meaning originate or the relation between language and action. Moreover progress in this area can allow us to develop autonomous artifacts able to cooperate and communicate to solve real-life problems.

Abstract:  This commentary considers some of the basic issues in the development of autonomous virtual agents from a rather general and theoretical viewpoint. It is predicated on an understanding of agents as ideal Bayesian observers, which follows in the long tradition of Helmholtzian ideas about how the brain works and subsequent developments in machine learning and computational neuroscience. The aim of this commentary is to define some key aspects of the problem and discuss potential solutions in relation to a series of specific research questions. In what follows, we try to cast the problem in terms of optimisation, which is particularly pertinent from the point of view of evolutionary schemes. The focus will be on maximizing the evidence for an agent’s model of his world or, more precisely, minimizing a variational free energy bound on negative model log-evidence or surprise. This has proven to be a useful framework in the computational neurosciences; and can be shown to be a fairly universal explanation for action and perception. Within this setting, the notion of a model (entailed by a subject) and a model of that subject (entailed by an agent observing the subject) is central. Framing the problem in terms of models raises key questions about their nature; particularly their dynamical form and implicit state spaces. A model-based perspective provides many clear answers to these questions. However, there are some key choices that may need to be formulated carefully, particularly in relation to difference between simply modelling the behaviour of a subject and modelling that behaviour under the constraint that the subject is modelling his world. We will focus on this distinction in terms of the difference between Bayesian and meta-Bayesian modelling of subject behaviours by virtual agents.

Abstract:  It is widely believed that the continuous change of neuronal connections provides animals with the capability to adapt and learn from experience. Several algorithms have been developed over the past 50 years to reproduce specific learning capabilities in artificial neural networks. However, little is known about the way in which these algorithms work together in a behaviorally-autonomous individual facing partially-unknown and changing environments. I will describe a series of experiments on the evolution of learning in autonomous robots that show the power of this adaptive method, provide an explanation to biological questions, and challenge assumptions on the link between synaptic plasticity and learning.

Abstract:  We postulate that a requirement for the development of autonomous agents capable of context dependent information processing entails that the dynamics of their neural network controllers are scale free. This idea is elaborated at two levels: the cognitive (mechanistic) and the phenomenal. At the cognitive level, this requirement is used as a constraint to limit the search space in the evolution of these agents. However, this constraint also has ramifications for the development of agents that exhibit a naturalized phenomenology in whom a neural correlate of subjective experience is identified. At this level, a scale free dynamics is a necessary condition for a subject to be able to interpret its own experience.

Abstract:  It is common place to say that the internal dynamical structure of autonomous agents evolved by biological principles should exist at “the edge of chaos.” In the present talk, we will try to address this view through the study of the main dynamical properties of a concrete evolvable agent. To describe the dynamical properties of an agent’s neural structure or “brain,” we will use a general paradigm developed using the qualitative theory of dynamical systems. This methodology explores the fundamental dynamical properties and patterns that an evolved complex agent should exhibit – focusing on the structure and patterns of attractors that reside within an evolved autonomous agent’s neural structure. We approach this analysis from a twofold perspective: on one hand, studying the interaction between the “brain,” the environment, and the tasks; on the other hand, the evolutionary process involved in the development of the autonomous agent. We will try to demonstrate how this analysis may aid in the evolution of autonomous agents by limiting the size of the search space.

Abstract:  Embodiment has long been identified as a central challenge in Artificial Intelligence. Robotic creatures interacting with the real world face the difficulties, and opportunities, afforded by the complexity, variability, and noise of the physical realm. We argue that robotics is not the only form of embodiment. Virtual agents living in online domains may interact with reality as well, with such interactions mediated by the physical creatures or “users” living in it. We will describe our experiments using a game to demonstrate co-adaptive phenomena between a collective of evolutionary online agents and their human users in the general context of co-evolutionary dynamics and investigate other cases of agent/human co-evolution.