Goal recognition in complex domains

In the paper, a technique for recognition of the goal of an observed action is presented. This technique is specifically suitable for domains which do not have a good domain theory but have an abundant set of episodes. An episode is composed of an action with the context in which it was performed and an expert´s interpretation of its goal. This technique uses the action´s past goals to define the goal space to determine the present goal and its associated belief. The similarity between the contexts of the previous executions of the action and the current context is used to determine the probability of each of the goals in the goal space. The goal with the highest probability is recognized as the goal of the action. The degree of belief of this goal is computed as a function of the entropy, relative entropy, and the cardinality of the determined probability distribution