Coevolutionary learning is a flexible paradigm with many applications, but its practice is hindered by various subtle pathologies. The Discovery of Objectives via Clustering (DOC) algorithm is a heuristic approach for learner selection that employs the maximization of expected utility (MEU) solution concept. DOC shows some potential for addressing certain issues; however, modification is found necessary to prevent the pathology of overspecialization. We propose QueMEU, a novel test generator and memory mechanism that uses a sampling policy over a queue to maintain diversity and provide an effective learning gradient. Incorporation of Que- MEU improves the performance of DOC on abstract numbers game problems as well as the challenging density classification task. QueMEU outperforms the standard test generator as well as a more established alternative that employs fitness sharing.

Co-evolution is a powerful problem-solving approach. However, fitness evaluation in co-evolutionary algorithms can be computationally expensive, as the quality of an individual in one population is defined by its interactions with many (or all) members of one or more other populations. To accelerate co-evolutionary systems, we introduce phylogenyinformed interaction estimation, which uses runtime phylogenetic analysis to estimate interaction outcomes between individuals based on how their relatives performed against each other. We test our interaction estimation method with three distinct co-evolutionary systems: two systems focused on measuring problem-solving success and one focused on measuring evolutionary open-endedness. We find that phylogeny-informed estimation can substantially reduce the computation required to solve problems, particularly at the beginning of long-term evolutionary runs. Additionally, we find that our estimation method initially jump-starts the evolution of neural complexity in our open-ended domain, but estimation-free systems eventually “catch-up” if given enough time. Further refinements to these phylogeny-informed interaction estimation methods offer a promising path to reducing the computational cost of running co-evolutionary systems while maintaining their open-endedness.

Much activity in the field of artificial life has been concerned with the search for appropriate and effective methods of quantifying open-ended behavior in evolutionary systems. The MODES Toolbox is a recent addition that employs a persistence filter over evolutionary lineages to focus attention only on those genotypes that are most adaptive. The Toolbox provides a useful and intuitive set of metrics in terms of change, novelty, complexity, and ecology. One domain thought to exhibit open-ended dynamics, the linguistic prediction game, is a ripe candidate for deeper statistical analysis. We apply the MODES Toolbox on this domain, lending support to prior hypotheses regarding evolutionary stable states while also suggesting that, in at least one case, the observed complexity growth is largely nonadaptive.

We study the conditions in which the unbounded growth of complexity – measured in terms of expressed genome size – can be observed in coevolving populations of neural agents involved in different classes of interactions. To reproduce the results of prior work on the dynamics of open-ended evolution, we introduce a simple pursuit-evasion scenario that allows for the development of increasingly intricate strategies. It is shown that for some configurations of our game, fitness-proportionate selection leads to stagnation while more sophisticated coevolutionary methods produce apparently unbounded complexity growth. Analysis of behavioral patterns sheds some light on the evolutionary pressures introduced by the model. Our findings replicate many features of previously reported work; however, we observe particular dynamics that differ in important respects, challenging prior conclusions, creating new opportunities, and highlighting the need for further investigation of this domain.