"Goal Recognition Design - Artificial intelligence techniques in surveillance problems”

This thesis focuses on Goal Recognition, an innovative and still developing field of research in Artificial Intelligence, which involves determining an agent's goal by observing its behaviour. Several real-world applications can be modelled in this context, for example, surveillance problems, operations where it is necessary to preserve user privacy, chat-boxes, and human-robot interaction. In this dissertation, we consider an environment in which an intelligent agent, the target, performs actions that can be observed by another intelligent agent, the observer (or by the environment itself). We analyze an evasive or fleeing target that moves in the environment to reach its destination, while the observer tries to find it. An example of this scenario is a surveillance drone that flies over a wide geographical area to discover a criminal who is trying to reach a hideout by car as soon as possible. In this context, the observer runs Goal Recognition algorithms that aim to discover where the target is going. In particular, the observer needs to reason about the target's strategy to obscure its goals. To support this kind of reasoning, we propose a technique to calculate a new measure, the "undisclosing index", that represents the maximal length of the prefix of a path that a fleeing agent may take before its goal becomes apparent to the observer. Moreover, we present methods for the observer to reason about how the target might change its behaviour based on the resources that it has available: the target will take shorter or longer paths based on his total travel budget. We modelled the environment in which the agents move by using a connected graph. For our experiments, we construct the graphs based on data from a real web map provided by OpenStreetMap (OSM). We implemented a web application designed for this purpose. Through it, the user selects and exports an area of interest on the map, identifying the target's possible goals within that area. The web-app provides a viewable version of the graph, implements the algorithms that help the observer reason about the possible paths of the target based on its budget and visualize them. The experimental evaluation shows how the solutions proposed in this thesis are useful tools for solving Goal Recognition problems.