The goal/intention recognition problem is the task of identifying an agent’s intent by observing its behaviour. Traditionally, the problem has involved matching a sequence of observations to a plan in a pre-defined plan library; the winning plan being the one that “best” matches the observations.
Recent developments dispense with the overhead of a plan library and instead—based on the assumption that the observed agent is behaving rationally—take a cost-based approach and uses classical planning technology to generate candidate plans as needed over a domain model.
In a series of work, we have extended and improved the cost-based approach to goal recognition, both for general task planning and path planning. We have worked on how to make recognition faster and more robust to irrational/erratic observed behavior. We have also looked at its’ dual problem: deceptive behavior, the problem of generating behavior that is as deceptive as possible, while still achieving the intended objective.
Some representative papers are:
- Peta Masters, Sebastian Sardiña: Expecting the unexpected: Goal recognition for rational and irrational agents. Artificial Intelligence. 297: 103490 (2021)
- Artem Polyvyanyy, Zihang Su, Nir Lipovetzky, Sebastian Sardiña: Goal Recognition Using Off-The-Shelf Process Mining Techniques. AAMAS 2020: 1072-1080
- Peta Masters, Sebastian Sardiña: Cost-Based Goal Recognition in Navigational Domains. Journal Artificial Intelligence Research 64: 197-242 (2019)
- Peta Masters, Sebastian Sardiña: Deceptive Path-Planning. IJCAI 2017: 4368-4375.
Sebastian Sardina
Professor in Artificial Intelligence
My research falls in the intersection between knowledge representation for reasoning about dynamic systems (reasoning about action and change), automated planning and reactive synthesis, and agent-oriented programming.