Search

Aiming at the prediction of truck travel time in open pit mines, we established a prediction model based on long short-term memory(LSTM). This model fully accounts for 11 factors, including the nature of trucks, weather, road conditions, and driver’s behaviors, as well as the influence of neighbor road segments in the route on the current predicted road segment. The experiment shows that the error of the LSTM prediction model is significantly reduced compared with SVR and BP models. In addition, the maximum absolute mean error under different conditions is less than 12 seconds.

This paper proposes a general method for the visualization fitness landscapes. The method transforms the original fitness landscape to a 3D visualized network from the perspective of the optimization algorithms. This method can visualize the fitness landscapes of continuous problems of different dimensions, combinatorial problems and even the mixed integer problems. Experiments show that the proposed method can maintain most of the critical properties of problems, such as modality, ruggedness and neutrality.

With increasing people who suffer from diet-related diseases, providing suggestions for personal daily nutrient-dense intake is highly expected. However, current dietary nutrition models are less precise, and dietary nutrition optimizers usually fail to give satisfactory solutions. Therefore, we construct a constrained many-objective nutrition model with more precise nutrient assessments and a scalable constrained many-objective benchmark set. This test suite has great flexibility in evaluating algorithms’ performance on high dimensional search and objective spaces with some feasible region fragments. We also propose a kd-tree based dynamic constrained many-objective evolutionary algorithm to search for customized food combinations according to personal daily consumption and intake preference. Experi ments show that our algorithm has better diversity maintenance ability in high dimension space.

Environmental selection is an important process in multi-objective evolutionary algorithms (MOEAs). As the evolution progresses, the number of non-dominated solutions increases. This paper is focused on selecting a subset from excess non-dominated solutions for evolutionary or the final output. However, traditional selection methods in classical MOEAs encounter difficulties when dealing with candidate solutions that possess irregular topologies. Although the distance-based subset selection methods are not sensitive to the topologies of the candidate points, they have significant room for reducing computational complexity. In order to address the above issues, a subspace selection method is proposed in this paper. It partitions the objective space into multiple subspaces that have comparable volumes and shapes. The maximal minimum distance of each solution is considered to ensure that the sparsest solution is always chosen first. To save computational costs, only the solutions in the neighboring subspaces are taken into account. The experimental results demonstrate that the proposed subspace selection method outperforms classical selection methods in solving problems with various shapes of the Pareto front.