Beam Search is a search algorithm that explores a graph by expanding the most promising nodes in a limited set, rather than considering all possible nodes. It is commonly used in traitement du langage naturel and traduction automatique tâches, où l'objectif est de générer des séquences, telles que des phrases ou des expressions.
Unlike traditional search algorithms that may evaluate all possible paths (like breadth-first or recherche en profondeur), Beam Search maintains a fixed number of the best candidates at each step, known as the largeur du faisceau. This means that only the top ‘N’ most promising candidates are kept for further exploration, while the rest are discarded. This approach helps to reduce the computational complexity and improves the efficiency of the search process.
Par exemple, dans un de langage generating text, at each word prediction step, Beam Search would consider the top ‘N’ most likely words based on previous context, forming a new set of candidate sentences. The algorithm continues this process until a complete sentence is formed, selecting the best options at each stage.
One important aspect of Beam Search is that it balances between exploration (trying out various options) and exploitation (focusing on the best-known options). A smaller beam width may lead to faster computations but risks missing better solutions, while a larger beam width may provide more comprehensive results at the cost of increased ressources informatiques.
En pratique, la recherche par faisceau est largement utilisée dans des applications telles que reconnaissance vocale, text summarization, and chatbot development, where generating coherent and contextually appropriate outputs is crucial.