Intention Recognition in the Situation Calculus and Probability Theory Frameworks

A method to recognize agent’s intentions is presented in a framework that combines the logic of Situation Calculus and Probability Theory. The method is restricted to contexts where the agent only performs procedures in a given library of procedures, and where the system that intends to recognize the agent’s intentions has a complete knowledge of the actions performed by the agent. An original aspect is that the procedures are defined for human agents and not for artificial agents. The consequence is that the procedures may offer the possibility to do any kind of actions between two given actions, and they also may forbid to perform some specific actions. Then, the problem is different and more complex than the standard problem of plan recognition. To select the procedures that partially match the observations we consider the procedures that have the greatest estimated probability. This estimation is based on the application of Bayes’ theorem and on specific heuristics. These heuristics depend on the history and not just on the last observation. A PROLOG prototype of the presented method has been implemented.