Evolving agents to learn how to solve complex, multi-stage tasks to achieve a goal is a challenging problem. Problems such as the River Crossing Task are used to explore how these agents evolve and what they learn, but it is still often difficult to explain why agents behave in the way they do. We present the Minimal River Crossing (RC-) Task testbed, designed to reduce the complexity of the original River Crossing Task while keeping its essential components, such that the fundamental learning challenges it presents can be understood in more detail. Specifically to illustrate this, we demonstrate that the RC- environment can be used to investigate the effect that a cost to movement has on agent evolution and learning, and more importantly that the findings obtained as a result can be generalised back to the original River Crossing Task.

Understanding how brains and environments give rise to behavior is a subject of great multidisciplinary interest. C. elegans is well-suited for this work because of its relatively rich behavioral repertoire and tractable connectome. The chemotaxis of C. elegans is comprised of two complimentary strategies - “weathervane” (klinotaxis) and “pirouette” (klinokinesis) - that operate in parallel with one another. The present work seeks to characterize each strategy and its contribution to the overall chemotaxis behavior. We find that the contribution of klinotaxis is the primary contributor of chemotaxis performance in most environments, but that klinokinesis is effective in environments with faint stimuli, have few gradient sources or are noisy, particularly when it is integrating sensed concentration over a longer time-scale.

We implement Varela, Maturana and Uribe's original autopoiesis algorithm with suitable modifications as proposed by McMullin. We further investigate how environmental factors affect formation of autopoietic entities - namely how long an entity remains a whole after formation and to what size does it grow in its life span i.e. Life span and Cycle size respectively. We find that ratios of different basic elements like Holes, Substrates and Catalysts do not affect Life span and Cycle size meaningfully but both properties are affected negatively if disintegration probability—the probability of a Link element to transform into a Substrate element—is increased.

Artificial life originated and has long studied the topic of open-ended evolution , which seeks the principles underlying artificial systems that innovate continually, inspired by biological evolution. Recently, interest has grown within the broader field of AI in a generalization of open-ended evolution, here called open-ended search , wherein such questions of open-endedness are explored for advancing AI, whatever the nature of the underlying search algorithm (e.g. evolutionary or gradient-based). For example, open-ended search might design new architectures for neural networks, new reinforcement learning algorithms, or most ambitiously, aim at designing artificial general intelligence. This paper proposes that open-ended evolution and artificial life have much to contribute towards the understanding of open-ended AI, focusing here in particular on the safety of open-ended search. The idea is that AI systems are increasingly applied in the real world, often producing unintended harms in the process, which motivates the growing field of AI safety. This paper argues that open-ended AI has its own safety challenges, in particular, whether the creativity of open-ended systems can be productively and predictably controlled. This paper explains how unique safety problems manifest in open-ended search, and suggests concrete contributions and research questions to explore them. The hope is to inspire progress towards creative, useful, and safe open-ended search algorithms.

The trade-off between the number and closeness of friendships is one of the important features of communication systems. This distinguishes asynchronous text communication through the Internet (lightweight communications) from the face-to-face communication and social grooming of primates (elaborate communications). In this study, we modeled communication as messaging flows driven by edge and node memory mechanisms in order to investigate micro-mechanisms that realize the trade-off law and the differences between the two types of communications. Five patterns of social structures including the trade-off law emerged depending on the strengths of the memory mechanisms. This suggests how communication systems construct different social structures. These results provide insight into the design of online social networks.

Catastrophic forgetting continues to severely restrict the learnability of controllers suitable for multiple task environments. Efforts to combat catastrophic forgetting reported in the literature to date have focused on how control systems can be updated more rapidly, hastening their adjustment from good initial settings to new environments, or more circumspectly, suppressing their ability to overfit to any one environment. When using robots, the environment includes the robot's own body, its shape and material properties, and how its actuators and sensors are distributed along its mechanical structure. Here we demonstrate for the first time how one such design decision (sensor placement) can alter the landscape of the loss function itself, either expanding or shrinking the weight manifolds containing suitable controllers for each individual task, thus increasing or decreasing their probability of overlap across tasks, and thus reducing or inducing the potential for catastrophic forgetting.

We provide conceptual clues for one promising Artificial Life (ALife) route to Artificial Intelligence (AI) based on the notion of habit. We draw from an enactive approach that considers habits as the building blocks for mental life and, consequently, as the foundation for a science of mind. By taking this standpoint, this approach departs from the conventional view of intelligence in AI, which is based on “higher-order” cognitive functions. The first part of the paper addresses the idea of taking habits as the foundation for modeling intelligent behavior. This requires us to consider the so-called “scaling up” problem and rethink the concept of intelligence that still pervades in mainstream cognitive science. In the second part, we present the enactive approach to habits, emphasizing their adaptive and complex nature, as well as their fundamental role in guiding behavior. Finally, we acknowledge some limitations in the current enactive models of habits: either they are disembodied and decoupled, but allow for a rich landscape of attractors, or they are embodied and coupled, but remain too minimal. We propose a bridge between existing models and point to the need to go beyond the individual to include a social domain. We conclude that to better model intelligent behavior, embodied and situated agents must be capable of developing an increasingly complex network of habits from which an intelligent self emerges.

Computer architectures that presume global hardware determinism are ultimately unscalable, but they are relatively easy to program because each operation is strictly sequenced and has an assured effect. Architectures that forgo global determinism can be indefinitely scalable , but they demand a shift in programming concepts, towards software mechanisms that can perform useful work given only limited, local synchronization and merely best-effort determinism. This paper introduces a parallel computing framework called SPOT— Stage Priority Operation Teams —designed for ulam programmers coding for the 2D grid of the Movable Feast Machine. Spatial threads , a simple but flexible 1D distributed programming paradigm, are introduced as a first use-case for SPOTs, with sample tasks ranging from moving objects to search, quorum sensing, and data reductions. SPOT and spatial threads also make intriguingly literal connections between crass physical concepts—such as space, time, and motion—and computational concepts such as program and data, and suggest a humble but fundamentally sensible meaning for the term ‘object’.

In domains where measures of utility for automatically-designed artefacts (or agents performing subjective tasks) are difficult or impossible to mathematically describe (such as ‘be interesting’), human interactive search algorithms are an attractive alternative. However, despite notable achievements, they are still designed around a specific search method, resulting in a lack of problem generality: applying a new search algorithm requires an excessive amount of redesign such that an altogether new interactive method is formed in the process. This leads to missed opportunities for human interactive methods to utilize the power of state of the art optimization algorithms. Here, we introduce for the first time a framework for human interactive optimization that is agnostic to both the search method and the application problem. Using 13 different search methods on 24 fitness functions commonly found in evolutionary algorithm benchmarks, we show that our approach works on the majority of tested applications: many of the search methods, provided with access to the fitness functions, performed no better than our framework, which employs surrogate human participants who act as less informed and erroneous representations of the fitness function. In this way, our framework for interactive optimization provides a scalable solution by facilitating the integration of numerous types of current state of the art or future search algorithms. Future work will involve generalizing this method to admit multi-objective optimization methods and validation with human participants.

Robot swarms can solve tasks that are impossible or too hazardous for single robots. For example, following a nuclear radiation leak, a user may wish to establish a distributed communication chain that partly extends into the most dangerous areas to gather new information. The challenge is to create long chains while maintaining chain connectivity (‘connected reach’), where those at the distant end of the chain are more likely to be disconnected. Here we take the concept of dynamic ‘boldness’ levels from animal behavior ( Stegodyphus social spiders) to explore such risky environments in a way that adapts to the size of the group. Boldness is implemented as a continuous variable associated with the risk appetite of individuals to explore regions more distant from a central base. We present a decentralized mechanism for robots, based on the frequency of their social interactions, to adaptively take on ‘bold’ and ‘shy’ behaviors. Using this new bioinspired algorithm, which we call SPIDER, swarms are shown to adapt rapidly to the loss of bold individuals by regenerating a suitable shy–bold distribution, with fewer bolder individuals in smaller groups. This allows them to dynamically trade-off the benefits and costs of long chains (information retrieval versus loss of robots) and demonstrates the particular advantage of this approach in hazardous or adversarial environments.

Computational scientists studying cognition, robotics, and Artificial Intelligence have discovered that variation is beneficial for many applications of problem-solving. With the addition of variation to a simple algorithm, local attractors may be avoided (breaking out of poor behaviors), generalizations discovered (leading to robustness), and exploration of new state spaces made. But exactly how much variation and where it should be applied is still difficult to generalize between implementations and problems as there is no guiding theory or broad understanding for why variation should help cognitive systems and in what contexts. Historically, computational scientists could look to biology for insights, in this case to understand variation and its effect on cognition. However, neuroscientists also struggle with explaining the variation observed in neural circuitry (neuronal variation) so cannot offer strong insights whether it originates externally, internally, or is merely the result of an incomplete neural model. Here, we show preliminary data suggesting that a small amount of internal variation is preferentially selected through evolution for problem domains where a balance of cognitive strategies must be used. This finding suggests an evolutionary explanation for the existence of and reason for internal neuronal variation, and lays the groundwork for understanding when and why to apply variation in Artificial Intelligences.

In spatial computational models such as cellular automata (CA), designing mobile objects larger than the CA neighborhood is challenging when the object properties and dynamics are incompletely specified in advance. This paper introduces C211 , a two-dimensional digital ‘protocell’ with life-like and potentially useful features, designed for the best-effort asynchronous CA called the Movable Feast Machine (MFM). The protocell consists of an amorphous variable-density ‘cytoplasm’ that uses gossiping to coordinate operations such as cell movements, surrounded by an asymmetric ‘bilayer membrane’ providing some environmental isolation while adapting to cytoplasmic dynamics. C211 was engineered in a new ‘little language’ called SPLAT, which adds discrete 2D spatial pattern transforms to the ulam programming language. SPLAT is expressive enough that minimal code was required, for example, to enable membrane topology changes such as cell splitting and fusion. C211 ’s cytoplasm maintains internal state but leaves dozens of bits unused per atom, while its membrane is purely stigmergic and stateless—so vast tracts of pristine CA state space remain available for future cellular dynamics, whether engineered, evolved, or both.

Fine-scale evolutionary dynamics can be challenging to tease out when focused on broad brush strokes of whole populations over long time spans. We propose a suite of diagnostic metrics that operate on lineages and phylogenies in digital evolution experiments with the aim of improving our capacity to quantitatively explore the nuances of evolutionary histories in digital evolution experiments. We present three types of lineage measurements: lineage length, mutation accumulation, and phenotypic volatility. Additionally, we suggest the adoption of four phylogeny measurements from biology: depth of the most-recent common ancestor, phylogenetic richness, phylogenetic divergence, and phylogenetic regularity. We demonstrate the use of each metric on a set of two-dimensional, real-valued optimization problems under a range of mutation rates and selection strengths, confirming our intuitions about what they can tell us about evolutionary dynamics.

Criticality is thought to be crucial for complex systems to adapt, at the boundary between regimes with different dynamics, where the system may transition from one phase to another. Numerous systems, from sandpiles to gene regulatory networks, to swarms and human brains, seem to work towards preserving a precarious balance right at their critical point. Understanding criticality therefore seems strongly related to a broad, fundamental theory for the physics of life as it could be, which still lacks a clear description of how it can arise and maintain itself in complex systems. In order to investigate this crucial question, we combine critical learning with evolutionary simulation for a population of Ising-embodied neural networks, striving to find resources distributed over a 2D environment. The results show compelling dynamics in the combination of critical learning with evolutionary computation, highlighting the exploratory nature of critical systems and the pragmatism of evolutionary algorithms. We also analyze the genotypic exploration strategy, exhibiting a tension between local and global scale adaptation.

In this paper, we apply the Polymerase-Exonuclease-Nickase Dynamic Network Assembly (PEN DNA) toolbox, a modular framework for molecular computing, to reservoir computing. While reservoir computing is traditionally implemented with recurrent neural networks, any system with similar recurrent properties, here chemical reaction networks (CRNs), can be used as a reservoir. We compared our approach to a previous CRN implementation of reservoir computing by Goudarzi et al. Our implementation yielded similar performance with respect to their benchmark tasks. We then took advantage of the modularity of the PEN DNA toolbox to investigate the impact of the CRN size on performance, both by hand and with an automated optimization process. In both cases, we were able to find systems with excellent performance while also being realistic with respect to in vitro implementation. Finally, we investigated the impact of constraining the weights of the output layer to be positive. This constraint guarantees that the system will remain relatively small, and thus makes it easier to implement in vitro . While this constraint led to an expected degradation in performance, we were still able to find good implementations of the reservoir.

We explore artificial spin ice (ASI) as a substrate for material computation . ASI consists of large numbers of nanomagnets arranged in a 2D lattice. Local interactions between the magnets gives rise to a range of complex collective behavior. The ferromagnets form large networks of nonlinear nodes, which in many ways resemble artificial neural networks. In this work, we investigate key computational properties of ASI through micromagnetic simulations. Our nanomagnetic system exhibits a large number of reachable stable states and a wide range of available dynamics when perturbed by an external magnetic field. Furthermore, we find that the system is able to store and process temporal input patterns. The emergent behavior is highly tunable by varying the parameters of the external field. Our findings highlight ASI as a very promising substrate for in-materio computation at the nanoscale.

Natural evolution and complex adaptations often surprise scientists. However, the creativity of evolution is not limited to the natural world, transcending any particular substrate. In the context of digital evolution, artificial organisms evolving in computational environments are also able to elicit surprise and wonder. Indeed, most digital evolution researchers can relate anecdotes highlighting how common it is for their algorithms to creatively subvert their expectations or intentions, expose unrecognized bugs in their code, produce unexpectedly potent adaptations, or engage in behaviors and outcomes uncannily convergent with ones found in nature. Such stories routinely reveal the surprise and creativity of evolution in these digital worlds, but they rarely fit into the standard scientific narrative and are treated as obstacles to be overcome rather than interesting results. Bugs are fixed, experiments are refocused, and one-off surprises become stories traded among researchers through lossy, inefficient and error-prone oral tradition. Moreover, to our knowledge, no collection of such anecdotes has been published before and many natural scientists do not recognize how lifelike digital organisms are and how natural their evolution can be. We have crowd-sourced the writing of a paper and collected first-hand reports from artificial life and evolutionary computation researchers, creating a written, fact-checked collection of entertaining and important stories. It serves to show that evolutionary surprise generalizes beyond the natural world, and may indeed be a universal property of all complex evolving systems.

Susceptibility to common human diseases such as cancer is influenced by many genetic and environmental factors that work together in a complex manner. The state-of-the-art is to perform a genome-wide association study (GWAS) that measures millions of single-nucleotide polymorphisms (SNPs) throughout the genome followed by a one-SNP-at-a-time statistical analysis to detect univariate associations. This approach has identified thousands of genetic risk factors for hundreds of diseases. However, the genetic risk factors detected have very small effect sizes and collectively explain very little of the overall heritability of the disease. Nonetheless, it is assumed that the genetic component of risk is due to many independent risk factors that contribute additively. The fact that many genetic risk factors with small effects can be detected is taken as evidence to support this notion. It is our working hypothesis that the genetic architecture of common diseases is partly driven by non-additive interactions. To test this hypothesis, we developed a heuristic simulation-based method for conducting thought experiments about the complexity of genetic architecture. We show that a genetic architecture driven by complex interactions is highly consistent with the magnitude and distribution of univariate effects seen in real data. We compare our results with measures of univariate and interactions effects from two large-scale GWAS studies of sporadic breast cancer and find evidence to support our hypothesis that is consistent with the results of our thought experiment.

Using a glider in the Game of Life cellular automaton as a toy model, we explore how questions of origins might be approached from the perspective of autopoiesis. Specifically, we examine how the density of gliders evolves over time from random initial conditions and then develop a statistical mechanics of gliders that explains this time evolution in terms of the processes of glider creation, persistence and destruction that underlie it.