Landscape Autoencoders (LAE): Learning the Geometry of Combinatorial Optimization

Landscape Autoencoders (LAE): Learning the Geometry of Combinatorial Optimization

Landscape Autoencoders (LAE) are a novel approach I developed for exploring difficult combinatorial optimization problems, initially motivated by research into Ramsey Theory.

Traditional optimization algorithms search through enormous discrete solution spaces, often with little understanding of their underlying structure. LAEs attempt to learn that structure by training an autoencoder to preferentially reconstruct solutions with desirable mathematical properties, effectively learning a representation of the problem's high-quality solution landscape.

The key innovation is a modified autoencoder loss function that rewards both reconstruction accuracy and solution quality. Rather than minimizing only the conventional reconstruction error, LAE minimizes a ratio of the form L = MSE(x, x̂) / (Fitness(x̂) + ε), where Fitness measures how well the reconstructed candidate satisfies the target mathematical objective. For Ramsey Theory, fitness can measure the quality of graph colorings that avoid forbidden monochromatic cliques structured as cross-entropy loss. This creates an incentive for the learned representation to favor promising regions of the solution space rather than simply reconstructing the input distribution.

I have implemented an initial version of LAE for Ramsey graph exploration. The broader hypothesis is that the learned latent space may capture useful geometric relationships among high-quality solutions, enabling more efficient generation, interpolation, and discovery of new candidates. Although initially designed for Ramsey Theory, the approach could potentially generalize to other combinatorial optimization problems and mathematical searches where candidate solutions can be represented numerically and evaluated with a fitness function. Further experiments and implementation details will be published here.