Video recording of the Open-Ended Evolution Workshop (http://www.alife.org/ws/oee1) held at the European Conference on Artificial Life (http://ecal2015.alife.org) at the University of York, UK, on 20 July 2015.
ECAL 2015
Part 2/5: Tim Taylor (part 2, start time 0:00), Guillaume Beslon (start time 2:44), Dave Ackley (start time 13:21).
Video recording of the Open-Ended Evolution Workshop (http://www.alife.org/ws/oee1) held at the European Conference on Artificial Life (http://ecal2015.alife.org) at the University of York, UK, on 20 July 2015.
Video recording of the Open-Ended Evolution Workshop (http://www.alife.org/ws/oee1) held at the European Conference on Artificial Life (http://ecal2015.alife.org) at the University of York, UK, on 20 July 2015.
updated 9 years ago
Video recording of the Open-Ended Evolution Workshop (http://www.alife.org/ws/oee1) held at the European Conference on Artificial Life (http://ecal2015.alife.org) at the University of York, UK, on 20 July 2015.
The ecosystem of web applications faces a critical paradox: on one hand, the Internet is a constantly evolving and unpredictable computing platform, on the other hand, the software services that run on top of it hardly have the ability to adapt to the evolution of this platform. Among the software services, we distinguish between service providers that provide micro services and service consumers that aggregate several micro services to deliver macro services to customers. Providers and consumers must handle uncertainty: providers cannot know in advance what consumers need; consumers rely on third-parties that can disappear at any time. Our proposal analogizes the software consumer / provider network to a bipartite ecological graph. This analogy provides the foundations for the design of EVOSERV, an individual-based ALife simulator used to experiment with decentralized adaptation strategies for providers and consumers. The initial model of a software network is tuned according to observations gathered from real-world software networks. The key insights about our experiments are that, 1) we can successfully model software systems as an ALife system, and 2) we succeed in emerging a global property from local decisions: when consumers and providers adapt with local decision strategies, the global robustness of the network increases. We show that these results hold with different initial situations, different scales and different topological constraints on the network.
Peripersonal space refers to the area around the body that is perceived as secure and reachable. The ability to build such a representation is necessary in both approach and avoidance behaviors. Several studies show that the perception of reachable and comfort areas depends on emotions. In this paper, we describe how we model an appetitive and an aversive pathway based on the role of some brain regions. The obtained emotional states modulate the robot perception of its peripersonal space. This representation is directly used to control the robot behavior. Based on a single-resource multirobot experiment, we show the impact of such an emotional modulation. Aggressive or fearful behaviors emerge from the dynamics of interaction between the simulated robots.
Living systems are organised in space. This imposes constraints on both their structural form and, consequently, their dynamics. While artificial life research has demonstrated that embedding an adaptive system in space tends to have a significant impact on its behaviour, we do not yet have a full account of the relevance of spatiality to living self-organisation.
Here, we extend the REDS model of spatial networks with self-organised community structure to include the “small world” effect. We demonstrate that REDS networks can become small worlds with the introduction of a small amount of random rewiring. We then explore how this rewiring influences two simple dynamic processes representing the contagious spread of infection or information.
We show that epidemic outbreaks arise more easily and spread faster on REDS networks compared to standard random geometric graphs (RGGs). Outbreaks spread even faster on small world REDS networks (due to their shorter path lengths) but initially find it more difficult to establish themselves (due to their reduced community structure). Overall, we find that small world REDS networks, with their combination of short characteristic path length, positive assortativity, strong community structure and high clustering, are more susceptible to a range of contagion dynamics than RGGs, and that they offer a useful abstract model for studying dynamics on spatially organised living systems.
Particle Swarm Optimisation (PSO) is a metaheuristic used to solve search tasks and is inspired by the flocking behaviour of birds. Traditionally careful tuning of parameters are required to avoid stagnation. Many animals forage using search strategies that show power law distributions in their motions in the form of Le ́vy flight random walks. It might be expected that when exploring spaces for optima in the absence of any prior knowledge a similar strategy may be useful. Using feedback from swarm metrics, we engineer modifications to the standard PSO algorithm that induce criticality. Such dynamics show long tail distributions in system event sizes. The presence of large (though few) exploratory steps removes the risk of stagnation. The Critical Particle Swarm (CriPS) can be easily combined with many existing PSO extensions.
Cooperation is observed widely in nature and is thought an essential component of many evolutionary processes, yet the mechanisms by which it arises and persists are still unclear. Among several theories, network reciprocity — a model of inhomogeneous social interactions — has been proposed as an enabling mechanism to explain the emergence of cooperation. Existing evolutionary models of this mechanism have tended to focus on highly heterogeneous (scale-free) networks, hence typically assume preferential attachment mechanisms, and consequently the prerequisite that individuals have global network knowledge. Within an evolutionary game theoretic context, using the weak prisoner’s dilemma as a metaphor for cooperation, we present a minimal model which describes network growth by chronological random addition of new nodes, combined with regular attrition of less fit members of the population. Specifically our model does not require that agents have access to global information and does not assume scale-free network structure or a preferential attachment mechanism. Further our model supports the emergence of cooperation from initially non-cooperative populations. By reducing dependency on a number of assumptions, this model offers broad applicability and as such may support an explanation of the emergence of cooperation in early evolutionary transitions, where few assumptions can be made.
In the past, evolved virtual creatures (EVCs) have been developed with rigid, articulated bodies, and with soft bodies, but never before with a combination of the two. In nature, however, creatures combining a rigid skeleton and non-rigid muscles are some of the most complex and successful examples of life on earth. Now, for the first time, creatures with fully evolved rigid-body skeletons and soft-body muscles can be developed in the virtual world, as well. By exploiting and re-purposing the capabilities of existing soft-body simulation systems, we can evolve complex and effective simulated muscles, able to drive a rigid-body skeleton. In this way, we can begin to bridge the gap between articulated and softbodied EVCs, and take the next step on a nature-inspired path to meaningful morphological complexity for evolved virtual creatures.
Collective decision making is crucial in human organizations and societies. When a collective is working on exploration of problem space and/or ideation for creative solutions, the evolutionary perspective is useful for conceptualizing and modeling collective decision making, where populations of ideas spread and evolve on a social network habitat via continual applications of evolutionary operators by human individuals (Sayama & Dionne 2015).
Using an evolutionary approach to model collective decision making, we conducted agent-based simulations to investigate how collective decision making would be affected by the size and topology of social network structure (Sayama, Dionne & Yammarino 2013). In our model, each agent has its own utility function defined over a multi-dimensional problem space, which is marginally different from the “true” utility function that is not accessible from any agent. Each agent can memorize multiple ideas in mind, and iteratively applies evolutionary operators (e.g., replication, mutation, recombination, elimination) to the idea population it has. The outcomes of evolutionary operations are stored in the agent’s mind, and also sent to the neighbors to which the agent is connected. This allows the spread of ideas through social ties.
Each simulation was run for a fixed number of iterations, and then the level of idea convergence at a social level and the true utility value of the most supported idea were measured as the performance metrics of collective decision making. The former metric characterizes the ability for the society to form consensus, while the latter characterizes its ability to find the true best solution. The size of the network (number of agents or nodes) was varied from 5 to 640 logarithmically. Three network topologies with the same average node degree were tested: Random, small-world and scale-free. For more details of the model and the simulation settings, see Sayama, Dionne & Yammarino (2013).
Simulation results indicated that expanding the size of the social network generally improved the quality of ideas at the cost of decision convergence. Simulations with different social network topologies further revealed that collective decision making on small-world networks with high local clustering tended to achieve highest decision quality more often than on random or scale-free networks (Fig. 1). This can be understood in that local clustering helps agents in different regions in a network maintain their respective focus areas and engage in different local search, possibly enhancing the effective parallelism of collective decision making and therefore resulting in a greater number of successful decisions. In contrast, the links in random and scale-free networks are all “global”, mixing discussions prematurely and therefore reducing the effective parallelism of collective decision making. These observations have an interesting contrast with the fact that random and scale-free networks are highly efficient in information dissemination because of their global connectedness. Our results indicate that such efficiency of information dissemination may not necessarily imply the same for collective decision making. This work may also offer an evolutionary explanation for the nontrivial relationship between network clustering and problem solving, a problem that is actively debated in organizational science (Shore, Bernstein & Lazer 2014).
The most prominent property of life on Earth is its ability to evolve. It is often taken for granted that self-replication– the characteristic that makes life possible–implies evolvability, but many examples such as the lack of evolvability in computer viruses seem to challenge this view. Is evolvability itself a property that needs to evolve, or is it automatically present within any chemistry that supports sequences that can evolve in principle? Here, we study evolvability in the digital life system Avida, where self-replicating sequences written by hand are used to seed evolutionary experiments. We use 170 self-replicators that we found in a search through 3 billion randomly generated sequences (at three different sequence lengths) to study the evolvability of generic rather than hand-designed self-replicators. We find that most can evolve but some are evolutionarily sterile. From this limited data set we are led to conclude that evolvability is a likely–but not a guaranteed– property of random replicators in a digital chemistry.
Recombination is ubiquitous in multicellular plants, animals and even fungi. Many studies have shown that recombination can generate a great amount of genetic innovations, but it is also believed to damage well-adapted lineages, causing debates over how organisms cope with such disruptions. Using an established model of artificial gene regulatory networks, here we show that recombination may not be as destructive as expected. Provided only that there is selection for developmental stability, recombination can establish and maintain lineages with reliably better phenotypes compared to asexual reproduction. Contrary to expectation, this does not appear to be a simple side effect of higher levels of variation. A simple model of the underlying dynamics demonstrates a surprisingly high robustness in these lineages against the disruption caused by recombination. Contrary to expectation, lineages subject to recombination are less likely to produce offspring suffering truncation selection for instability than asexual lineages subject to simple mutation. These findings indicate the fundamental differences between recombination and high mutation rates, which has important implications for understanding both biological innovation and hierarchically structured models of machine learning.
We consider whether selection for evolvability leads to greater adaptive progress than selection for adaptedness alone. Our treatment bears on longstanding discussions of selection for evolvability in the literature, which have been largely limited to conceptual and qualitative arguments to date. We study a simple mathematical model of a population of individuals whose adaptedness and evolvability (here modelled as the standard deviation of mutations affecting adaptedness) are both under selective forces. In the special case of a population of size two, we show that the optimal amount of selection for evolvability depends on the ratio between the initial evolvability and the amount that evolvability can increase in the time given. Our result shows that to maximize the amount of adaptation it never pays off to select for evolvability more than to select for adaptedness itself. We have not answered the question of to what degree evolvability is selected for in nature, however we have made a small step in quantitative modelling of the evolution of evolvability and proved the existence of conditions under which selection for evolvability has a demonstrably positive effect.
In this paper, we investigate the closed-loop acquisition of basic behaviours on Sphero – a real spherical robot. We impose the additional requirement of learning from scratch in a single episode. The behaviour is encoded as an inverse model for stabilization and sensory target tracking tasks using recurrent neural networks.
This paper describes a set of simulations in which autonomous robots are required to coordinate their actions in order to transport a cuboid object that is too heavy to be moved by single robot. Robots’ controllers are synthesised using artificial evolution and dynamic neural networks. We compare two different types of robots: in the NT-condition, the robots are equipped with a camera and proximity sensors. In the T-condition, the robots have additional torque sensors. The result shows that best evolved groups of the T-condition outperform those of the NT-condition. Moreover, we show that the best evolved groups can adapt to variability in size and weight of the object as well as to the small variability in the cardinality of the group. We also show that simple forms of recruitment behaviour emerges without being selected for during evolution. This work unveils interesting relationships between design choices and characteristics of the evolved solutions, and it contributes to develop design guidelines for engineering robust and successful swarm robotic systems.
DNA is not the sole medium by which parents transmit information to their offspring. Epigenetic inheritance, in particular, is based on the partial transmission of the cellular state of the parental cell to its descendants. Although the reality of epigenetic inheritance is now firmly established, whether it has an influence on the long term evolutionary process is still subject to debate.
To address this question, we used RAevol, an in silico experimental evolution platform, and defined 4 scenarios with static or dynamic environments and with or without epigenetic inheritance. Simulations in dynamic environments show that protein inheritance indeed increases the rate of evolution on the long term. But they also show that it impedes evolution in its very first stages. This negative effect can be explained by instabilities generated by the interference between the two inheritance mediums. On the opposite, the long term gain can be explained by protein inheritance reducing the constraints on the genetic regulation network.
Edge detection is a fundamental procedure in image processing, machine vision, and computer vision. Its application area ranges from astronomy to medicine in which isolating the objects of interest in the image is of a significant importance. However, performing edge detection is a non-trivial task for which a large number of techniques have been proposed to solve it. This paper investigates the use of Ant Colony Optimization — a prominent set of optimization heuristics — to solve the edge detection problem. We propose two modified versions of the algorithm Ant Colony System (ACS) for an efficient and a noise-free edge detection.
Iterated learning takes place when the input into a particular individual’s learning process is itself the output of another individual’s learning process. This is an important feature to capture when investigating human language change, or the dynamics of culturally learned behaviours in general. Over the last fifteen years, the Iterated Learning Model (ILM) has been used to shed light on how the population-level characteristics of learned communication arise. However, until now each iteration of the model has tended to feature a single immature language user learning from their interactions with a single mature language user. Here, the ILM is extended to include a population of immature and mature language users. We demonstrate that the structure and make-up of this population influences the dynamics of language change that occur over generational time. In particular, we show that, by increasing the number of trainers from which an agent learns, the agent in question learns a fully compositional language at a much faster rate, and with less training data. It is also shown that, so long as the number of mature agents is large enough, this finding holds even if a learner’s trainers include other agents that do not yet posses full linguistic competence.
In this paper we address the automatic synthesis of controllers for the coordinated movement of multiple mobile robots. We use a noise-resistant version of Particle Swarm Optimization to learn in simulation a set of 50 weights of a plastic artificial neural network. Two learning strategies are applied: homogeneous centralized learning, in which every robot runs the same controller and the performance is evaluated externally with a global metric, and heterogeneous distributed learning, in which robots run different controllers and the performance is evaluated independently on each robot with a local metric. The two sets of metrics enforce Reynolds’ flocking rules, resulting in a good correspondence between the metrics and the flocking behaviors obtained. Results demonstrate that it is possible to learn the collective task using both learning approaches. The solutions from the centralized learning have higher fitness and lower standard deviation than those learned in a distributed manner. We test the learned controllers in real robot experiments and also show in simulation the performance of the controllers with increasing number of robots.
Tierra is an iconic ALife system. It has three innovative features: self-optimisation, parasitism, and an uncrashable execution environment. We have identified four sources of bias in the evolutionary dynamics of Tierra. We have run two sets of simulations: one with the original configuration of Tierra 6.02, and one with three of the biases removed. We find that the innovations observed in the original Tierra are preserved, and the debiased configuration is more amenable for future experiments with open-ended evolution.
Emergence is a phenomenon taken for granted in science but also still not well understood. We have developed a model of artificial genetic evolution intended to allow for emergence on genetic, population and social levels. We present the details of the current state of our environment, agent, and reproductive models. In developing our models we have relied on a principle of using non-linear systems to model as many systems as possible including mutation and recombination, gene-environment interaction, agent metabolism, agent survival, resource gathering and sexual reproduction. In this paper we review the genetic dynamics that have emerged in our system including genotype-phenotype divergence, genetic drift, pseudogenes, and gene duplication. We conclude that emergence-focused design in complex system simulation is necessary to reproduce the multilevel emergence seen in the natural world.
In this paper, we extend our previous model circuit for steering in C. elegans to control a more realistic biomechanical model of forward locomotion. We show that the identified steering circuit is sufficient to steer the full body during forward locomotion while only innervating a few of the anterior most neck muscles. Analysis of the sensorimotor transformation and phasic stimulation experiments provides evidence that the principles of operation for steering discussed in the model are relevant for steering in the worm. Finally, the integration of the steering circuit in a physical model of the full body allows us to compare more closely the properties of the evolved solutions with those of the worm.
As more is becoming understood about how the brain represents and computes with high-level spatial information, the prospect of constructing biologically-inspired robot controllers using these spatial representations has become apparent. Grid cells are particularly interesting in this regard, as they provide a general coordinate system of space. Artificial neural network models of grid cells show the ability to perform path integration, but important for a robot is also the ability to calculate the direction from the current location, as indicated by the path integrator, to a remembered goal. Present models for goal-directed navigation using grid cells have used a simulating approach, where the networks are required to actively test successive locations along linear trajectories emanating from the current location. This paper presents a passive model, where differences between multiscale grid cell representations of the present location and the goal are used to calculate a goal-direction signal directly. The model successfully guides a simulated agent to its goal, showing promise for implementing the system on a real robot in the future. Some possible implications for neuroscientific studies on the goal-direction signal in the entorhinal/subicular region are briefly discussed.
Embodied cognition is the hypothesis that behavior is not simply caused by the brain. Instead, behavior emerges from the interactions between brains in particular kinds of bodies embedded in environments that provide certain kinds of opportunities for activity. Theories of embodied cognition require a mechanism to support how these distributed resources can remain in contact with one another so that they can be assembled into task-specific solutions to problems. These theories of embodied cognition, especially the nonrepresentational kinds, typically rely on James J. Gibson’s (1979) notion of ecological information as the relevant mechanism, and there is extensive empirical support for the claim that this information both exists and is used by organisms to coordinate and control their activity.
For Gibson, information refers to structures in ambient energy arrays (e.g. light for vision) that are specific to the object or event in the world that caused the structure. This structure becomes information when we use it to coordinate and control our behavior, and this information supports what Gibson (1979) referred to as the direct perception of our environments.
This account of how we maintain psychological contact with the world has been formalized in the years since Gibson’s death in the context of task dynamics (e.g. Saltzman & Kelso, 1987). Objects and events in the world can only be identified uniquely at the level of dynamics (the description of how a system changes over time, with reference to the forces causing the change; Bingham, 1995). Some of these dynamics can then be mapped into energy arrays as kinematic patterns (descriptions of change with no reference to underlying forces). These kinematic patterns, although not identical to the dynamical objects or events, can be specific to them (Runeson & Frykholm, 1983; Turvey, Shaw, Reed & Mace, 1981) and therefore be informative about them. Detecting a specifying kinematic pattern is equivalent to perceiving the underlying dynamic, and our perceptual systems are exquisitely sensitive to these patterns.
A given task dynamic can, by definition, only produce information about that particular dynamic. Information-based control of behavior is therefore task-specific, and so empirical research that investigates information-based explanations for behavior therefore follows four crucial steps (Wilson & Golonka, 2013; see there for specific research examples): 1. Identifying the task facing the organism at the level of task dynamics; 2. Use this task dynamical analysis to identify a comprehensive list of the specific resources offered by the task to support a solution, in particular the task-specific information variables created by the task dynamics; 3. Describe how these resources are assembled into a solution to the task at hand; 4. Test the model and it’s predictions in real organisms
The purpose of this presentation is to introduce this ecological, biological notion of information and the related research programme to the artificial life community, where ‘information’ typically refers to measures of entropy, following Shannon. An information theoretic analysis may prove useful to our scientific understanding of ecological information but regardless, that is not the kind of information biological systems interact with as they perceive and act in their environments. That is Gibson’s information. Understanding this ecological information is a critical part of understanding how biological life gets up to the things that it does. It may also, therefore, be of use to helping artificial life get up to those same things.
Animals demonstrate a level of agility currently unmatched in their robotic counterparts. The elasticity of muscles and tendons increase not only performance, but also the efficiency of movements. In contrast, robots are often constructed with rigid components connected by motors. However, recently compliant actuators and materials have been introduced to enhance robot designs, emulating the flexibility of natural organisms. In this paper, we incorporate passive flexibility into the spine of a quadruped animat and employ computational evolution to generate gaits. Results indicate that spine flexibility significantly increases both performance and efficiency of evolved individuals. Moreover, evolving the degree of spine flexibility along with artificial neural network controllers produces the highest performing solutions.
Environmental factors that determine ecological niches, for example natural boundaries formed by mountains, rivers, deserts, contribute to the speciation among animals. Similar factors have been proposed to be important for the emergence of cultural and technical innovations in human populations in the pre-state stages of societies. Here we describe a social simulation aimed to investigate this issue. The simulation uses two environmental features, mountain ridges and the fertility of the land. The results show that indeed these environmental factors matter for the emergence of successful innovative populations. The defenses provided by mountain ridges facilitate the emergence of many populations with moderately successful innovations. The fertile lands are where the populations with the most successful innovations emerge, however in most cases these populations can trace their origins to innovative populations emerging under the defense of mountain ridges. This simulation study provides experimental support for the relatively speculative theories about the importance of environmental factors for the emergence of cultural and technical innovations.
This paper investigates the dynamics of a simple coevolutionary system. It consists of a predator-prey system in which one population maximizes its distance to the members of the other population, while the second population tries to minimize the distance to the first population. This results in a coevolution ary pursuer-evader (PE) system whose dynamics can easily be studied. Next, a simple genotype-phenotype mapping is added to the system. This mapping as well as other sources of increased selection push the system towards regions of maximum adaptability (ROMAs). These ROMAs are a generalization of the concept ”evolution to the edge of chaos”.
Complexity science often uses generative models to study and explain the emergent behavior of humans, human culture, and human patterns of social organization. In spite of this, little is known about how the lowest levels of human social organization came into being. That is, little is known about how the earliest members of our hominini tribe transitioned from being presumably small-groups of ape-like polygamous/promiscuous individuals (beginning perhaps as early as Ardipithecus or Australopithecus after the time of the Pan-Homo split in the late Pliocene to early Pleistocene eras) into family units having stable breeding-bonds, extended families, and clans. What were the causal mechanisms (biological, possibly cognitive, social, and environmental, etc.) that were responsible for the conversion? To confound the issue, it is also possible the conversion process itself was a complex system replete with input sensitivities and path dependencies i.e., a nested complex system. These processes and their distinctive social arrangements may be referred to favorably (as one notable anthropologist has called them) as, “the deep structure of society.” This essay describes applied research that uses discrete event computer modeling techniques in an attempt to model-then-understand a few of the underlying social, environmental, and biological systems present at the root of human sociality; at the root of social complexity.
Video recording of the Open-Ended Evolution Workshop (http://www.alife.org/ws/oee1) held at the European Conference on Artificial Life (http://ecal2015.alife.org) at the University of York, UK, on 20 July 2015.
Video recording of the Open-Ended Evolution Workshop (http://www.alife.org/ws/oee1) held at the European Conference on Artificial Life (http://ecal2015.alife.org) at the University of York, UK, on 20 July 2015.
Video recording of the Open-Ended Evolution Workshop (http://www.alife.org/ws/oee1) held at the European Conference on Artificial Life (http://ecal2015.alife.org) at the University of York, UK, on 20 July 2015.
Video recording of the Open-Ended Evolution Workshop (http://www.alife.org/ws/oee1) held at the European Conference on Artificial Life (http://ecal2015.alife.org) at the University of York, UK, on 20 July 2015.