Lexicase selection is an effective many-objective evolutionary algorithm across many problem domains. Lexicase can be computationally expensive, especially in areas like evolutionary robotics where individual objectives might require their own physics simulation. Improving the efficiency of Lexicase selection can reduce the total number of evaluations thereby lowering computational overhead. Here, we introduce a fitness agnostic adaptive objective sampling algorithm using the filtering efficacy of objectives to adjust their frequency of occurrence as a selector. In a set of binary genome maximization tasks modeled to emulate evolutionary robotics situations, we show that performance can be maintained while computational efficiency increases as compared to ϵ -Lexicase.

Neural networks are often chosen as controllers in evolutionary robotics. In all but a few cases, neural networks are evolved from scratch. In this study, we investigate the effect of pretraining neural networks using a biologically inspired walking gait. We first generate joint angles for a walking gait using an inverse kinematics model. We then train a conventional feed-forward neural network to reproduce these joint angles. The pretrained model is used to seed an initial population of neural networks, which are coevolved along with the morphology of a quadrupedal robot using Lexicase selection. Our initial results show that while pretraining does not necessarily lead to higher fitness at the end of evolution, it does lead to more consistent performance and more lifelike final behaviors. This exploration has left us with many questions about the importance and process of pretraining in evolutionary robotics, and we believe our results suggest the technique is worth further investigation.

Neural networks (NNs) are effective controllers for evolutionary robotics, imposing few limits on potential gaits. Morphology evolved with a controller enables brain and body to become tightly coupled. Typically, NN parameters (sometimes architectures) and animat bodies are randomly initialized at the start of evolution. In this paper, we pretrain NNs with supervised learning, bootstrapping NN outputs towards oscillating behaviors prior to evolution. We focus on quadrupedal gaits as they are well-studied in biology and several common gait patterns have been identified, named, and studied by the research community. We hypothesize that performance of evolved gaits will improve with pretraining compared to beginning evolution with randomly initialized NNs. Our results show that only some pretraining regimens outperform (in terms of distance traveled and viability) random initialization of NN parameters. Furthermore, some regimens introduce an initial bias that is difficult to overcome, resulting in better initial performance but worse performance in the long term.

Controllers capable of exhibiting multiple behaviors is a longstanding goal in artificial life. Evolutionary robotics approaches have demonstrated effective optimization of robotic controllers, realizing single behaviors in a variety of domains. However, evolving multiple behaviors in one controller remains an outstanding challenge. Many objective selection algorithms are a potential solution as they are capable of optimizing across tens or hundreds of objectives. In this study, we use Lexicase selection evolving animats capable of both wall crossing and turn/seek behaviors. Our investigation focuses on the objective sampling strategy during selection to balance performance across the two primary tasks. Results show that the sampling strategy does not significantly alter performance, but the number of evaluations required varies significantly across strategies.

Digital simulation enables a wide variety of research and applications underlying the study of artificial life. In evolutionary robotics applications, the focus is often on maximizing performance of an animat for a specific task. Analyzing evolved behaviors can be challenging, however, given the complex coupling of morphology and brain. In this paper, we introduce a simulation environment built to investigate animats capable of smoothly transitioning between operating modes (e.g., from cautious to aggressive or from one physical form to another). The simulator provides functionality for logging sensory information as well as animat state enabling a deep analysis. Although more abstract than soft-body or rigid-body physics engines, it is lightweight and efficient, allowing for a high number of simulations in a small amount of time. The simulation supplements other more complex physics-based environments providing for greater inspection of sensor information and animat behavior. Furthermore, it is designed to provide an extensible test bed beyond just gait transitions to assess new artificial intelligence and evolutionary algorithms and more importantly the combination of these techniques.

Generalized behavior is a long standing goal for evolutionary robotics. Behaviors for a given task should be robust to perturbation and capable of operating across a variety of environments. We have previously shown that Lexicase selection evolves high-performing individuals in a semi-generalized wall crossing task–i.e., where the task is broadly the same, but there is variation between individual instances. Further work has identified effective parameter values for Lexicase selection in this domain but other factors affecting and explaining performance remain to be identified. In this paper, we expand our prior investigations, examining populations over evolutionary time exploring other factors that might lead to generalized behavior. Results show that genomic clusters do not correspond to performance, indicating that clusters of specialists do not form within the population. While early individuals gain a foothold in the selection process by specializing on a few wall heights, successful populations are ultimately pressured towards generalized behavior. Finally, we find that this transition from specialists to generalists also leads to an increase in tiebreaks, a mechanism in Lexicase, during selection providing a metric to assess the performance of individual replicates.

Animals interact with their environment softly through interaction of muscles, tendons, and rigid skeleton. By incorporating flexibility, they reduce ground impact forces and improve locomotive efficiency. Flexibility is also beneficial for robotic systems, although it remains challenging to implement. In this paper, we explore the addition of passive flexibility to a quadrupedal animat; we measure the impact of flexibility on both locomotive performance and energy efficiency of movement. Results show that spine and lower limb flexibility can significantly increase distance traveled when compared to an animat with no flexibility. However, replacing passively flexibile joints with actively controlled joints evolves more effective individuals with similar efficiency. Given these results, the number of joints and joint configuration appear to drive performance increases rather than just the addition of passive flexibility.

Agents exhibiting generalized control are capable of solving a theme of related tasks, rather than a specific instance. Here, generalized control pertains to the locomotive capacity of quadrupedal animats, evaluated when climbing over walls of varying height to reach a target. In prior work, we showed that Lexicase selection is more effective than other evolutionary algorithms for this wall crossing task. Generalized controllers capable of crossing the majority of wall heights are discovered, even though Lexicase selection does not sample all possible environments per generation. In this work, we further constrain environmental sampling during evolution, examining the resilience of Lexicase to the impoverished conditions. Through restricting the range of samples at given points in time as well as fixing environmental exposure over fractions of evolutionary time, we attempt to increase the ‘adjacency’ of environmental samples, and report on the response of the Lexicase algorithm to the pressure of this reduced environmental diversity. Results indicate that Lexicase is robust, producing viable agents even in considerably challenging conditions. We also see a positive correlation between the number of tiebreak events that occur and the success of individuals in a population, except in the most limiting conditions. We argue that the increased number of tiebreaks is a response to local maxima, and the increased diversity resulting from random selection at this point, is a key driver of the resilience of the Lexicase algorithm. We also show that in extreme cases, this relationship breaks down. We conclude that tiebreaking is an important control mechanism in Lexicase operation, and that the breakdown in performance observed in extreme conditions indicates an inability of the tiebreak mechanism to function effectively where population diversity is unable to reflect environmental diversity.

A primary goal of evolutionary robotics (ER) is generalized control. That is, a robot controller should be capable of solving a variety of tasks in a domain, rather than only addressing specific instances of a task. Prior work has shown that Lexicase selection is more effective than other evolutionary algorithms for a wall crossing task domain where quadrupedal animats are evaluated on walls of varying height. In this work we expand baseline treatments in this task domain and examine specific aspects of the Lexicase selection algorithm across a variety of different parameter configurations. We identify the most effective Lexicase parameters for this task. Results indicate that Lexicase’s success is potentially due to maintaining population diversity at a higher level than other algorithms explored for this domain.

Evolving robust behaviors for robots has proven to be a challenging problem. Determining how to optimize behavior for a specific instance, while also realizing behaviors that generalize to variations on the problem often requires highly customized algorithms and problem-specific tuning of the evolutionary platform. Algorithms that can realize robust, generalized behavior without this customization are therefore highly desirable. In this paper, we examine the Lexicase selection algorithm as a possible general algorithm for a wall crossing robot task. Previous work has resulted in specialized strategies to evolve robust behaviors for this task. Here, we show that Lexicase selection is not only competitive with these strategies but after parameter tuning, actually exceeds the performance of the specialized algorithms.