The main objective of the project is to provide agents (both human and software) with recommendations on future behavior (actions, plans or goals) based on learning models of their past behavior and recognition of their present actions, plans or goals. In order to achieve the objective, we will integrate techniques from machine learning, activity/plan recognition and automated planning.
The key idea is to analyze information from sensors in order to generate diverse models of an agent’s behavior, such as planning domain models, probabilistic models, or generic classifiers. These models will allow the recognition system to use diverse methods to predict the plan/action/goal the agent is pursuing and build plans or provide recommendations that assist humans or software/robotic agents in their daily activities. This is a novel approach in relation to the current literature that focuses on at most two of these three aspects. Also, we plan to use a hybrid approach within each one of the three tasks by using and combining different existing and new techniques.
- Development of techniques to automatically synthesize behavioural models from past observations, executions and/or interactions.
- Development of planning recognition techniques to predict the actions or goal of an agent in the behavioural models from current observations.
- Development of reasoning mechanisms that leverage the result of the planning recognition to propose a new course of action, suggest new goals or recommend better-utility outcomes taking into account the agent’s preferences.
- Propose an architecture that allows engineers to rapidly develop plan recognition systems.
- Propose several testing scenarios as proof-of-concept in the area of plan and activity recognition.