Concurrent Metareasoning

Abstract

Metaresoning is again under focus in the AI community. Here in this paper, a new classification for types of metareasoning has been proposed. In recent years, only the ones that are here named as pre-metareasoning and para-metareasoning have been studied. The first one is for predicting the best computation path for having better performance programs. The second, mostly known as interruptible anytime algorithm, is to limit the computation time externally when the approximate answer is better than nothing. One other type of metareasoning (called here as post-metareasoning) is discussed in a case study. It has been shown as an effective method for reducing error in self-localization. Based on the measurements in the case study, the post-metareasoning argued as useful when the effectiveness of reasoning methods are not known by the designer or when the system learns the reasoning methods and should evaluate and use the best one automatically. As the post-metareasoning is based on the results of different isolated reasoning methods, it is possible to be handled in parallel. The speed of post-metareasoning in such a case is determined by the time required by the slowest reasoning method and the post-metaresoning itself.