Research On Algorithm Recognition Based On Autonomous Mental Development

Algorithm Recognition is considered with program understanding. It is an important but difficult problem. In the past decades, several approaches have been studied in this area, but most of them are based on a library where predefined templates are stored. The templates are composed of program statements that implement the algorithms. The process of such algorithm recognition is actually programtemplate matching. Sometimes, this approach works well, but such template-based approaches encounter an obstacle that it is difficult to know how many templates are required to recognize a given algorithm in advance.Such uncontrolled variation of an algorithm can be regarded as one of muddy characteristics of a real task. To address the muddiness, an approach called Autonomous Mental Development has been proposed. The advantage of the Autonomous Mental Development approach is that it can works without predefining the task it may deal with. In this paper, we explore the possibility that the AMD approach could be applied to algorithm recognition. We suggest employing the Autonomous Mental Development approach to "raise" the templates for algorithm recognition instead of to predefine the templates. Thus, determining the number of implementation versions of an algorithm in advance can be avoided.However, the Autonomous Mental Development approach, such as Lobe Component Analysis, is suitable for metric vector spaces while program codes are non-vectorial items. For this reason, a vector space model for program codes is proposed fisrt. It is used to convert each program code into a vector so that the Autonomous Mental Development approach could be applied.Then, we give a detailed discussion of the developmental templates for algorithm recognition. The experiment results demonstrate that our approach is feasible and reach 93.4% recognition accuracy in average. Last we give the conclusion and possible future directions for this work.