Search Algorithm Using Prior Knowledge of Similar Task

— Evaluation of Search Ability Using Toy Problem —

Abstract

This study introduces a concept of transfer learning to search tasks such as function approximation and optimization, and aims to achieve effective search using fewer number of samples. We propose a search procedure transfer (SPT) algorithm which extracts knowledge of efficient search procedure from well-known task and uses it to search similar unknown tasks. Experiments reveal that the closer the target task becomes to the source task, the higher the performance of the SPT algorithm becomes. In addition, experiments also demonstrate the SPT algorithm shows higher performance than the existing method using the Bayesian optimization algorithm (BOA) when the difference between a source task and a target task is small. Applying the idea of domain randomization, the SPT algorithm can utilize even human ambiguous heuristic knowledge.