Planning is the model-based approach to autonomous behavior in AI. Given a compact description of a model representing the actions, the sensors, and goals of an agent, a planner must automatically produce the control for driving the agent to its goal. The approach contrasts with model-free approaches, where the agent behavior results from learning, or programming-based approaches, where the agent behavior is hardwired.
In this course, we will look at the variety of models used in AI planning for representing actions, sensors, and goals, and the techniques that have been developed for solving these models. In the simplest of these models, actions have deterministic effects and the initial state of the environment is known (classical planning); in the most complex ones, actions have stochastic effects and sensing is partial and noisy (POMDP planning). All the models are intractable in the worst case, and hence, the key challenge in planning is computational: how to scale up to solve large models. This is a research oriented course, where we will review the state of the art in AI planning, what has been accomplished so far, and what problems are still open. Prior knowledge of LTL and model checking is desired but not essential.
The course will consist of around 20 hours (Category B: Internal PhD courses) to be held at DIS, in Via Ariosto 25 with the following calendar:
Mon. 19 July, 10.00 -- 13.00, room A4 (aula A4)
Tue. 20 July, 10.00 -- 13.00, room A3 (aula A3)
Wed. 21 July, 10.00 -- 13.00, room A6 (aula A6)
Mon. 26 July, 10.00 -- 13.00, room A6 (aula A6)
Tue. 27 July, 10.00 -- 13.00, room A6 (aula A6)
Wed. 28 July, 10.00 -- 13.00, room A6 (aula A6)