In recent years, swimming robots have been developed to achieve efficient propulsion and high maneuverability that are possessed naturally by fish. Previous studies have attempted to achieve swimming similar to fish by control based on physical models and top-down architectures, but have encountered problems due to the high complexity of the underwater environment. Several research works have tried to overcome these problems by exploiting embodiment—that is, by mimicking the physical properties of fish. To achieve more intelligent swimming from the perspective of the embodiment, we focused on a framework called physical reservoir computing (PRC). This framework allows us to utilize physical dynamics as a computational resource. In this study, we propose a soft sheet-like swimming robot and a PRC-based architecture that can be used to emulate swimming motions by exploit- ing its own body dynamics for closed-loop control. Through experiments, we demonstrated that our system satisfies the properties required for learning swimming motion through supervised learning. We also succeeded in robust motion gen- eration and environmental state estimation, opening up future prospects for more intelligent robot control and sensing.
Direct reciprocity is one of the key mechanisms accounting for cooperation in our social life. According to recent under-standing, most of the classical strategies for direct reciprocity fall into one of two classes, ‘partners’ or ‘rivals.’ A ‘partner’ is a generous strategy achieving mutual cooperation, and a ‘rival’ never lets the co-player become better off. They have different working conditions: For example, partners show good performance in a large population, whereas rivals do in head-to-head matches. Using exhaustive enumeration us- ing a super-computer, we demonstrate the existence of strategies that are partners as well as rivals, called ‘friendly ri- vals.’ Among them, we focus on a human-interpretable strategy, named ‘CAPRI’ after its five characteristic ingredients, i.e., cooperate, accept, punish, recover, and defect otherwise. Our evolutionary simulation shows excellent performance of CAPRI regardless of environmental conditions.
We used a model of honeybees’ decision-making to investigate the effect of robots on the hive’s foraging decisions. We find that at least two robots are needed to make a significant impact.
On 26th February 2021, we premiered a theatre play with the title AI: When a Robot Writes a Play1 in the Švanda theatre in Prague, Czechia. The 60-minute play, followed by a 75-minute discussion, was streamed online via the webpage of the THEaiTRE project (Rosa et al., 2020), within which it was created.2 The play was performed in Czech and accompanied by English subtitles. The script (i.e., the dialogues) for the play was generated by our THEaiTRobot 1.0 system (Rosa et al., 2021) based on the GPT-2 language model (Radford et al., 2019) with very little human intervention. Specifically, our analysis (to be published soon) shows that 92% of the words in the script were generated by GPT-2. We describe some aspects of the generation process and some issues that we had to resolve. The subsequent staging of the play, on the other hand, was completely in the hands of human theatre professionals. The THEaiTRobot 1.0 system has several limitations, such as being capable of only generating scripts for individual scenes, not for a full play. We discuss some of the most important issues and our intended approach at solving them. This will eventually lead to a THEaiTRobot 2.0 system, which will be used to generate a script for a new theatre play, with a premiere planned for January 2022. The text of this extended abstract is largely based on the THEaiTRobot 1.0 short paper by the same authors (Rosa et al., 2021).
When we attempt to define life, we tend to refer to individuals, those that are alive. But these individuals might be cells, organisms, colonies... ecosystems? We can describe living systems at different scales. Which ones might be the best ones to describe different selves? I explore this question using concepts from information theory, ALife, and Buddhist philosophy. After brief introductions, I review the implications of changing the scale of observation, and how this affects our understanding of selves at different structural, temporal, and informational scales. The conclusion is that there is no ``best'' scale for a self, as this will depend on the scale at which decisions must be made. Different decisions, different scales.
Animals that live in social groups must interact in order to stay together and move collectively. By socializing a robot with a group of weakly electric fish, we aim at answering fundamental biological questions about the rules that govern social interactions and cause group members to coordinate their movements and come to joint decisions. African weakly electric fish communicate at night by emitting and perceiving short electrical current pulses, a process called electrocommunication. Our experiments have shown that it is only possible to integrate a robotic agent into a group of electric fish if it emits electric signals and engages in electro-communication. All other sensory cues, like visual appearance, can be neglected. For full acceptance as a conspecific by live fish, the robot must be able to interact with the animals. We hypothesize that the integration of fish and robot into a mixed society can succeed when the robot's electric signaling interaction is matched by locomotor interactions that are congruent with the behavioral relevance of electro-communication.
—Conventional design for robotics is based on the assumption that the robot should operate only in one given environment. As a result, often their skills are not transferable. Biological systems on the other hand are surprisingly versatile and robust. They exhibit remarkable adaptivity by placing more emphasis on adapting their morphology. Consequently, provid-ing robots with mechanisms to adapt their bodies (material properties and even removing/adding parts) could be a way to obtain more versatile and robust systems. In this paper we propose a novel method which uses genetic algorithms to evolve optimal adaptation rules for changing the bodies of soft robots. Instead of optimising the morphology directly, we optimise the rules that tell the robot how to adapt the body based on the feedback it receives when interacting with the environment. It uses a combination of local and global information to sculpt (i.e., change stiffness and remove body parts) the soft body to improve locomotion in different environments. We show that in some cases the same rule with the same starting morphology can lead to different, but beneficial morphologies in different environments, i.e., it can translate feedback from the different environments into different useful bodily changes. Furthermore, we demonstrate that some of the found rules are highly robust and are able to produce successful morphologies for a range of environments that haven’t been experienced during the optimisation process.
We evolve artificial agents to perform a simple tracking task in three conditions: one individual (Isolated Condition) and two joint action conditions with division of labor. The joint conditions differ by whether two agents switch complementary roles during the task (Generalist Condition) or always play the same role (Specialist Condition). At the end of evolutionary runs we calculate the agents’ neural complexity using Tononi-Sporns-Edelman (TSE) complexity measure which relates to Integrated Information Theory (IIT). We show that (1) division of labor with specialization leads to a level of neural complexity comparable to the complexity of perform- ing the same task alone, and that (2) both are lower than neural complexity when performing the task jointly with role switching. We further consider viewing collaborating agents as a single extended system and calculate its joint neural com- plexity. We demonstrate that contrary to our predictions, the same pattern of results, i.e., Generalists’ complexity being higher than Specialists’, holds also in this conceptualization.
Reservoir computing is a powerful computational framework that is particularly successful in time-series prediction tasks. It utilises a brain-inspired recurrent neural network and al- lows biologically plausible learning without backpropaga-tion. Reservoir computing has relied extensively on the self- organizing properties of biological spiking neural networks. We examine the ability of a rate-based network of neural os- cillators to take advantage of the self-organizing properties of synaptic plasticity. We show that such models can solve complex tasks and benefit from synaptic plasticity, which in-creases their performance and robustness. Our results further motivate the study of self-organizing biologically inspired computational models that do not exclusively rely on end- to-end training.
We implement a probabilistic version of the game of life on a quantum annealer in a very direct manner, demonstrating possible use cases for (currently very noisy) quantum annealing devices in the simulation of (usually very stochastic) artificial life and complex systems in general.
This paper aims at considering novel and practical applications of ALife techniques to design a co-creative social dynamics in an online and virtual space, which is becoming important because of the recent emergence of various types of online communication platforms due to outbreaks of COVID-19. Recently, spatial and online communication services, such as SpatialChat, have attracted more attention. Each participant is represented as an avatar or icon in a virtual 2D space. She can move it around in the space and listen to neighbors’ voices of which volume become louder as they get closer to her. However, the overall structure of communications tends to be deadlocked, which might make participants lose chances to communicate with many other people. We design and investigate a virtual agent, called “facilitator agent,” as a study towards realization of practical agents that facilitate novel and cooperative interactions in a spatial and online communication by giving human participants opportunities to communicate with many others cooperatively. We adopt a Social Particle Swarm (SPS) model to simulate group dynamics in this type of communication service. We assume several behavioral patterns of a facilitator agent with fixed game-theoretical strategies and several movement strategies. We discuss how incorporating a single facilitator agent into the space can increase novel and cooperative interactions in several behavioral settings of the facilitator agent. We also report on a preliminary experiment on designing a facilitator agent using a deep reinforcement learning technique.
This programmatic paper continues a series of works that we are dedicating to introduce a novel research program in AI, which we call Autopoietic SB-AI to indicate two basic elements of its procedural architecture. (1) The first element is the innovative methodological option of synthetically studying the cognitive domain based on the construction and experimental exploration of wetware –i.e., chemical – models of cognitive processes, using techniques defined in the field of Synthetic Biology (SB). (2) The second element is the theoretical option of developing SB models of cognitive processes based on the theory of autopoiesis. In our previous works we focused on the epistemological and theoretical groundings of Autopoietic SB-AI. In this contribution, after a general presentation of this research program, we introduce the SB technical framework that we are developing to orient Autopoietic SB-AI towards a twofold goal: building organizationally relevant wetware models of minimal biological-like systems (i.e., synthetic cells), and, on this basis, contributing to the scientific exploration of minimal cognition.
Arrangements of nanomagnets known as artificial spin ices show great potential for use in unconventional computation. The majority of exploratory work done in this area considers just a small handful of well studied geometries (nanomag-netic arrangements), and uses them as if they were a black box. Here we detail a novel representation of artificial spin ice geometries, which lends itself to the tuning and evolu- tionary search of geometries. Using our representation we present geometries tuned to exhibit a desired computational or meta-material property. This is the first example of such a search performed on artificial spin ice.
Brains are among the most complex evolved objects. In re-cent years we have seen an explosion in the development of artificial cognitive systems constructed in silico (i.e. digital brains). In fact, we are now capable of creating digital brains whose operation is so complex that they are ef-fectively black boxes (Castelvecchi, 2016; Gunning, 2017). Previous work (Marstaller et al., 2013; Hintze et al., 2018; Kirkpatrick and Hintze, 2019) has identified and expanded upon various information-theoretic measures that can shed light on the internal processes of digital brains. Here we introduce a new information-theoretic measure called Fragmentation ( F ) which can measure how fragmented information is in an a digital brain. To provide a example of the application of F we look at the evolutionary emergence of complexity. Questions regarding the evolution of complexity have been of interest for as long as evolution has been a theory (Gregory, 1935). Nature is responsible for the development of a massive array of complex organisms, each comprised of various organs and regulatory systems that are themselves complex (McShea and Brandon, 2010). It has been observed that complexity can evolve even when complexity itself is being selected against (Beslon et al., 2021). We conclude by using F to show a case of evolved complexity that results in coincidental encryption.
Within the context of the European Horizon 2020 project ACDC
Donald Hebb proposed in his 1949 book The Organization of Behavior that cell assemblies organized by temporally-asymmetric excitation form the basis of cognition. This basic idea has inspired a large body of research in neuroscience, and to a lesser extent in artificial intelligence. The modern manifestation of Hebb’s principle is Spike-Timing Dependent Plasticity (STDP), and though we have a large body of exper- imental work investigating STDP, there is still little understanding of how networks of spiking neurons organize them- selves into complex functional circuits, even though some progress has been made with models such as Liquid State Machines. Networks popular in artificial intelligence (e.g. MLPs) and in artificial life (e.g. CTRNNs) tend to eschew Hebb’s insight and use error-backpropagation by gradient descent, in the case of AI, or an a-temporal Hebbian learning rule based on the outer product of neural activities, in the case of AL. Both of these approaches have greater interpretabil- ity than Spiking Neural Networks (SNNs), but both lack the mechanism that Hebb claimed was fundamental to cognition. This paper proposes to use complex-valued neurons (CVNs) to address this limitation, simultaneously promoting biologi- cal interpretation and computational tractability. The CVNs encode the firing rate and spike-time of a spiking neuron in the magnitude and angle, respectively, of a complex number. We also introduce an unsupervised piecewise-linear STDP learning rule compatible with CVNs, which for brevity we call complex-valued STDP (CVSTDP). We demonstrate both learning through error-backpropagation, and the spontaneous formation and dissolution of cell assemblies via the CVSTDP rule.
A key challenge in AI is the development of algorithms that are capable of cooperative behavior in interactions involving multiple independent machines or individuals. Of particular interest are social dilemmas, which are situations that raise tension between an individual’s best choice and the desirable outcome in terms of the group. Although such scenarios have been studied increasingly within the AI community recently, there are still many open questions on which aspects drive cooperative behavior in a particular situation. Based on the in- sights from behavioral experiments that have suggested positive effects of penalty mechanisms towards cooperation, in this work we adopt the notion of penalties by enabling independent and adaptive agents to penalize others. To that end, we extend agents’ action spaces with penalty actions and define a negative real-valued punishment value. We utilize reinforcement learning to simulate a process of repeated interaction between independent agents, learning by means of trial-and-error. Our evaluation considers different two player social dilemmas, and the N-player Prisoner’s Dilemma with up to 128 independent agents, where we demonstrate that the proposed mechanism combined with decentralized learning significantly increases cooperation within all experiments.
Diverse dynamical systems, living or not, were recently (1) discussed with focus on whether their evolution rate was decreasing or increasing. The first type of dynamics optimizes a target on fast timescales, e.g. all spontaneous physical processes minimize their free energy and bacterial monocultures optimize fitness in a constant environment (2). Evolutionary progress can be viewed as overcoming a series of ever increasing ‘record sized’ dynamical barriers of system specific sort (3; 4). The second type, evolutionary expansion, is less common and not as well understood (5). An example is given in (1) based on a time series of British GDP values per person per year (6), a quantity whose trend has kept increasing faster than exponentially since the late Middle Ages. Gross National Product (GDP) per capita (7) is not directly linked to the abundance or scarcity of material resources but rather gauges economic activities carried out by means of the infrastructures and institutions that maintain a population with shared cultural values and social and technological know-how. GDP history yields therefore insights into human cultural evolution. A recent study (8), to which we refer for most details, defines ’wealth’ as the ability to generate the economic activities e.g. measured by per capita GDP and introduces a simple causal albeit strongly aggregated model assuming that the observed wealth growth is mainly driven by human collaborative efforts whose intensity itself increases with increasing wealth. As detailed in (8), finite time singularities are strongly supported by key data sets (9; 10) and rapid changes in the mechanisms producing them must occur to avert social disruption. The mechanism proposed in (8) to generate the singularity could be of general relevance and begs the question of how finite time singularities extracted from data analysis can predict major transitions of evolutionary processes. As an example, GDP per capita data (6) from the UK (blue dots) and the corresponding model generated trend (red hatched curve) are plotted vs. time in Fig. 1. The insert 100
We propose a neural network based architecture to infer which parameters are fundamental, and their values, for pro-ducing specific instances of spatial patterns formed through cell colony growth. The system is trained on variations of the same pattern to recognize features that characterize it. Fur- thermore, selecting important parameters within our study mainly focuses on the fact that cells communicate. We use two forms of this communication as fundamental in finding the parameter values: bacterial conjugation, and environmen- tal signals. The neural network is trained during 3000 epochs to identify the pattern class and specific parameter values needed to reproduce the desired pattern. These parameters are then inputted into a gro simulation to assess proximity to the original pattern. Our architecture achieved a 5% error upon pattern reproduction.
Open-ended novelty is one of the goals of ALife. This pro- vides challenges for analysis as the system evolves. We pro-vide definitions for several emergent properties, such as para- sitism and hypercycles, observed to emerge in an RNA world configuration of the Stringmol automata chemistry, and show how these can simultaneously be mathematically simple, capture the complexity of the processes, and be readily imple- mentable.
