Evolving Through the Looking Glass: Learning Improved Search Spaces with Variational Autoencoders

Nature has spent billions of years perfecting our genetic representations, making them evolvable and expressive. Generative machine learning offers a shortcut: learn an evolvable latent space with implicit biases towards better solutions. We present SOLVE: Search space Optimization with Latent Variable Evolution, which creates a dataset of solutions that satisfy extra problem criteria or heuristics, generates a new latent search space, and uses a genetic algorithm to search within this new space to find solutions that meet the overall objective. We investigate SOLVE on five sets of criteria designed to detrimentally affect the search space and explain how this approach can be easily extended as the problems become more complex. We show that, compared to an identical GA using a standard representation, SOLVE with its learned latent representation can meet extra criteria and find solutions with distance to optimal up to two orders of magnitude closer. We demonstrate that SOLVE achieves its results by creating better search spaces that focus on desirable regions, reduce discontinuities, and enable improved search by the genetic algorithm. Fig. 1.Search space Optimization with Latent Variable Evolution (SOLVE). An optimizer produces a dataset of random solutions satisfying an extra criterion (e.g., constraint or secondary objective). A variational autoencoder learns this dataset and produces a learned latent representation biased towards the desired region of the search space. This learned representation is then used by a genetic algorithm to find solutions that meet the objective and extra criterion together