In recent years, the field of deep learning has been propelled forward by increasingly complex and resource-intensive neural network models. Despite their impressive performance, their training consumes much energy, incurring significant environmental impacts. Even more so, when neural architecture search (NAS) methods train hundreds of thousands of models to benchmark optimization. We offer to create greener benchmarks using genetic algorithms. We leverage our previous CNNGen approach producing random CNN topologies based on our context-free grammar. We then use the NSGA-II genetic algorithm to create topologies balancing performance and energy consumption. We rely on machine-learning-based predictors that estimate candidates' performance and energy consumption to avoid training during generations, saving significant computational costs. This paper reports on our experiments and discusses further developments.
This paper extends previous work in predictive maintenance that implement a self-adaptive evolutionary strategy (SA - ES) for feature selection in multi-label anomaly detection tasks. We incorporate Shapley Additive exPlanations (SHAP), an explainable artificial intelligence (XAI) method, to evaluate the relevance and classification performance of features selected by the SA - ES, and compare their effectiveness against those with the highest absolute Shapley values. A comparative analysis is performed using three multi-label classifiers on a public predictive maintenance dataset, with evaluation conducted through 5 - fold cross-validation. Our findings validate the efficiency of the SA - ES in reducing feature dimensions and minimizing Hamming Loss. Additionally, we present insightful visualizations and interpretations for multi-label anomaly classification, facilitating the application of predictive maintenance in real-world industrial scenarios.
Selectivity can be understood as a causal property that determines variable degrees of response specificity. In minimal autonomous systems, given their recursive dynamics, these responses are directed, not towards an hypothetical representation of the environment, but towards its own future states. In this sense, although co-specified by environmental circumstances, system state transitions can be characterized as dynamically more or less driven by an intrinsic 'steering force' underpinned by the degree of freedom that the organized states can provide under the influence of different environmental circumstances. In this work we present the idea of (latent) agency as a proto-cognitive property already present in autonomous systems and that can be generally measured through information metrics. We test these ideas and formulations through toy-experiments in the Game of Life cellular automaton.
Animal dispersal is among the most impacted behaviours from habitat loss and fragmentation. Animals disperse in search of better conditions to ultimately increase their fitness. However, the dispersal process induces costs that can be translated as a loss in fecundity or as an increase in predation risks. In this study, we use an evolutionary multi-agent model to quantify how costs related to information acquisition shape information gathering strategies used in animal dispersal. We separate costs of private information, i.e., related to the direct acquisition of information sensed in the physical environment, from costs related to social information, i.e., derived from the observation of conspecifics behaviour. We find that a low mortality risk associated with dispersal movements selects non-informed dispersal in low-variability environments. However, high-variability environments select an information gathering strategy based on the exploration of multiple patches of habitat. A higher mortality risk causes drastic drops in population abundance and agents gather information from only one patch to limit their movements. Overall, private information is predominantly used in highvariability environments while social and private information are used equally in low-variability environments.
In practice, various industrial design problems have continuous, discrete, and categorical variables. Our objective is to manage efficiently these different types of variables within a surrogate-based optimization process. In this work, we propose to redefine the notion of distance between the possible values of a categorical variable (named "attributes"), through the concept of "affinity". The notion of affinities between attributes can be interpreted as a weighted relationship between attributes. These affinities are usually defined based on a physical intuition of the designer. Indeed, affinities are generally implicitly associated to the behavior of one or several outputs that behave(s) similarly for various attributes. In order to study the impact of the use of affinities, numerical results are presented on specific test problems coming from structural and mechanical design frameworks.
