Theory Of Misleading Action And Its Application In Partially Observable Plan Recognition

Plan recognition is an active research field of artificial intelligence. Since from the year of 1978 that Schmidt, Sridharan and Goodson firstly proposed plan recognition as a research problem, more and more researchers go deep into this field. The most standing out was that Kautz and Allen proposed a general model for plan recognition in 1986, and this model nearly concluded all of submission of plan recognition, and it's the first formal theory of plan recognition.In former plan recognition, we often assumed that the actions carried out by the observed agent is true actions, and only for the ultimately goal. Under such restriction, we must believe that the actions performed by the observed agent are all necessary foundations for recognize, and ignoring to any action is false to the result. This assumption has obvious malpractice to two kinds of recognition. For the keyhole recognition, the mistake operates of the observed agent would be the recognition basis of the observation agent. And for the adverse recognition, the observed agent and the observation agent are in hostile condition, and thus there are three kinds of actions performed by the observed agent. The first is the necessary action for the goal, and the second is the mistake action, then the third is the misleading action which can induce the other agent makes wrong judgment. So it is not reasonable that the three kinds of actions all as the recognition basis.We must adopt appropriate method to exclude mistake actions and misleading actions, so that we can get exact recognition result. This paper aims at the problem of misleading actions. It introduces the concept of reliability, correlativity and correlated action sequence etc. And provides the calculate method of reliability which include the calculation both in fully and partially observable plan recognition. The paper also gives the algorithm which can estimate whether the observed action is misleading action or not according to the calculation result. Proposing the Maximum Cardinality Assumptions, and the assumptions are different from the Minimum Cardinal Assumptions from Kautz, which delete root nodes blindly. The assumptions this paper proposed aim at select the correlative root nodes wholely. Building a new recognizer which can add unobservable actions and remove misleading actions. Misleading action is widely exist, so the research to this problem can make for enriching the theory of plan recognition, and it also has significant scientific value for widen the implication field of plan recognition. This algorithm makes the recognition result more accuracy. And it would be very useful in intrusion detection and network security.