Abstract-Legged robots show promise for complex mobility tasks, such as navigating rough terrain, but the design of their control software is both challenging and laborious. Traditional evolutionary algorithms can produce these controllers, but require manual decomposition or other problem simplification because conventionally-used direct encodings have trouble taking advantage of a problem's regularities and symmetries. Such active intervention is time consuming, limits the range of potential solutions, and requires the user to possess a deep understanding of the problem's structure. This paper demonstrates that HyperNEAT, a new and promising generative encoding for evolving neural networks, can evolve quadruped gaits without an engineer manually decomposing the problem. Analyses suggest that HyperNEAT is successful because it employs a generative encoding that can more easily reuse phenotypic modules. It is also one of the first neuroevolutionary algorithms that exploits a problem's geometric symmetries, which may aid its performance. We compare HyperNEAT to FT-NEAT, a direct encoding control, and find that HyperNEAT is able to evolve impressive quadruped gaits and vastly outperforms FT-NEAT. Comparative analyses reveal that HyperNEAT individuals are more holistically affected by genetic operators, resulting in better leg coordination. Overall, the results suggest that HyperNEAT is a powerful algorithm for evolving control systems for complex, yet regular, devices, such as robots.

HyperNEAT represents a class of neuroevolutionary algorithms that captures some of the power of natural development with a computationally efficient high-level abstraction of development. This class of algorithms is intended to provide many of the desirable properties produced in biological phenotypes by natural developmental processes, such as regularity, modularity and hierarchy. While it has been previously shown that HyperNEAT produces regular artificial neural network (ANN) phenotypes, in this paper we investigated the open question of whether HyperNEAT can produce modular ANNs. We conducted such research on problems where modularity should be beneficial, and found that HyperNEAT failed to generate modular ANNs. We then imposed modularity on HyperNEAT's phenotypes and its performance improved, demonstrating that modularity increases performance on this problem. We next tested two techniques to encourage modularity in HyperNEAT, but did not observe an increase in either modularity or performance. Finally, we conducted tests on a simpler problem that requires modularity and found that HyperNEAT was able to rapidly produce modular solutions that solved the problem. We therefore present the first documented case of HyperNEAT producing a modular phenotype, but our inability to encourage modularity on harder problems where modularity would have been beneficial suggests that more work is needed to increase the likelihood that HyperNEAT and similar algorithms produce modular ANNs in response to challenging, decomposable problems.

N early 150 years ago, Charles Darwin explained how evolution and natural selection transformed the earliest life forms into the rich panoply of life seen today. Scientists estimate this process has been at work on Earth for at least 3.5 billion years.But we remain at the dawn of evolution in another world: the world of computing. There, evolution helps humans solve complex problems in engineering and provides insight into the evolutionary process in nature. As computing power continues to increase, researchers and developers apply evolutionary algorithms to an ever-widening variety of problems. As the "Evolution in a Computer" sidebar shows, evolutionary computation methods such as genetic algorithms have already achieved considerable success, rivaling and surpassing human designers in problem domains as wide-ranging as flash memory sticks and aircraft wings.We are investigating how to harness the power of evolution to help construct better computer software. The increasing interaction between computing technology and the physical world motivates this work. Systems must adapt to their environment, compensate for failures, optimize performance, and protect themselves from attacks-all with minimal human intervention. 1,2To design robust and resilient computational systems, we can take inspiration from nature. Living organisms have an amazing ability to adapt to changing environments, both in the short term through phenotypic plasticity and in the longer term through Darwinian evolution. Indeed, no existing cybersystem rivals the complexity of Earth's biosphere, yet life on Earth has evolved to not only deal with this complexity but to thrive on it.Many researchers have studied how to use the characteristics of natural systems to design better computing systems. One approach mimics the behaviors of social insects and other species. However, while such biomimetic methods have shown promise in controlling fleets of unmanned robotic systems and in other applications, they can only codify behaviors observed in nature today. Purely biomimetic approaches seek to imitate the results of evolution, but they do not account for the process of natural selection that produced those behaviors.For example, we can design the control software on a microrobot so that it mimics certain behaviors found in ants. However, while the robot might possess some physical characteristics reminiscent of an ant, the differences vastly outnumber the similarities. On the other hand, if we had the ability to evolve the control software, taking into account the capabilities of the robot and the characteristics of its environment, new behaviors might emerge that more effectively control the robot. 3 DIGITAL PETRI DISHDigital evolution gives us this power, and we are investigating how it can aid us in designing robust computational systems. Digital evolution is a form of evolutionary computation in which self-replicating computerIn digital evolution, self-replicating computer programs-digital organisms-experience mutations and selective pressures, po...

For centuries it was thought that bacteria live asocial lives. However, recent discoveries show many species of bacteria communicate in order to perform tasks previously thought to be limited to multicellular organisms. Central to this capability is quorum sensing, whereby organisms detect cell density and use this information to trigger group behaviors. Quorum sensing is used by bacteria in the formation of biofilms, secretion of digestive enzymes and, in the case of pathogenic bacteria, release of toxins or other virulence factors. Indeed, methods to disrupt quorum sensing are currently being investigated as possible treatments for numerous diseases, including cystic fibrosis, epidemic cholera, and methicillin-resistant Staphylococcus aureus. In this paper we demonstrate the evolution of a quorum sensing behavior in populations of digital organisms. Specifically, we show that digital organisms are capable of evolving a strategy to collectively suppress self-replication, when the population density reaches a specific, evolved threshold. We present the evolved genome of an organism exhibiting this behavior and analyze the collective operation of this "algorithm." Finally, through a set of experiments we demonstrate that the behavior scales to populations up to 400 times larger than those in which the behavior evolved.

Increasingly, high-assurance software systems apply selfreconfiguration in order to satisfy changing functional and non-functional requirements. Most self-reconfiguration approaches identify a target system configuration to provide the desired system behavior, then apply a series of reconfiguration instructions to reach the desired target configuration. Collectively, these reconfiguration instructions define an adaptation path. Although multiple satisfying adaptation paths may exist, most self-reconfiguration approaches select adaptation paths based on a single criterion, such as minimizing reconfiguration cost. However, different adaptation paths may represent tradeoffs between reconfiguration costs and other criteria, such as performance and reliability. This paper introduces an evolutionary computationbased approach to automatically evolve adaptation paths that safely transition an executing system from its current configuration to its desired target configuration, while balancing tradeoffs between functional and non-functional requirements. The proposed approach can be applied both at design time to generate suites of adaptation paths, as well as at run time to evolve safe adaptation paths to handle changing system and environmental conditions. We demonstrate the effectiveness of this approach by applying it to the dynamic reconfiguration of a collection of remote data mirrors, with the goal of minimizing reconfiguration costs while maximizing reconfiguration performance and reliability.