Omelet's Approach

Generation

The Omelet solver processes input data for a given problem, utilizing an encoder to analyze its features, and a decoder to rapidly generate diverse solutions. This generative AI-based solver acts as a prior, identifying generalized patterns across various problem datasets and enabling fast solution generation for the target problem.

Improvement Search

Following this, the Neural Guided Improvement Search refines the solutions generated by the decoder, progressively enhancing their quality. The parallelizable search process allows simultaneous exploration across diverse initial solutions, leading to a wider set of optimized outcomes. The AI module further guides the search, determining how to adjust and improve specific parts of the solutions. The process is akin to OpenAI's "chain of thought" approach, where solutions are iteratively constructed through reasoning steps.

End-to-end Training

By generating strong initial solutions and iteratively improving them, we can build a high-quality dataset internally. This dataset is then used to retrain the generative solver and improvement search modules, continuously enhancing the performance of the Omelet solver.