Evolutionary algorithm with an autoencoder-based dimensionality reduction mechanism

Many modern applied optimization problems are formulated as black-box models, characterized by a lack of analytical information about the objective function and its properties. While Evolutionary Algorithms (EAs) are frequently applied here, their performance degrades in high-dimensional spaces with complex landscapes. Moreover, these algorithms require generating a large number of trial solutions for effective operation, which may be unattainable in applications with expensive objective function evaluation. To overcome these limitations, we introduce a novel EA framework centered on adaptive, autoencoder-based dimensionality reduction. The autoencoder is retrained dynamically during the search using evolving population data. The strategy relies on the parallel operation of two optimization algorithms: one works in the original solution space, and another operates in a latent space – a compact, nonlinear representation of the current population generated by the autoencoder. This design enables the algorithm to adapt to the problem's specific structure. We evaluated the method on standard benchmark problems, analyzing its convergence dynamics and sensitivity to the latent space size. Statistical significance of the results was assessed using the Wilcoxon rank-sum test. The L-SRTDE algorithm was used as the subpopulation optimization method. Numerical experiments demonstrate that the proposed algorithm enhances search space exploration in first optimization stages but, on average, is slightly less efficient than the base algorithm.