Planning and learning under uncertainty

Resumen

Automated Planning is the component of Artificial Intelligence that studies the computational process of synthesizing sets of actions whose execution achieves some given objectives. Research on Automated Planning has traditionally focused on solving theoretical problems in controlled environments. In such environments both, the current state of the environment and the outcome of actions, are completely known. The development of real planning applications during the last decade (planning fire extinction operations (Castillo et al., 2006), planning spacecraft activities (Nayak et al., 1999), planning emergency evacuation actions (Muñoz-Avila et al., 1999)) has evidenced that these two assumptions are not true in many real-world problems. The planning research community is aware of this issue and during the last years, it has multiplied its efforts to find new planning systems able to address these kinds of problems. All these efforts have created a new field in Automated Planning called planning under uncertainty. Nevertheless, the new systems suffer from two limitations. (1) They precise accurate action models, though the definition by hand of accurate action models is frequently very complex. (2) They present scalability problems due to the combinatorial explosion implied by the expressiveness of its action models. This thesis defines a new planning paradigm for building, in an efficient and scalable way, robust plans in domains with uncertainty though the action model is incomplete. The thesis is that, the integration of relational machine learning techniques with the planning and execution processes, allows to develop planning systems that automatically enrich their initial knowledge about the environment and therefore find more robust plans. An empirical evaluation illustrates these benefits in comparison with state-of-the-art probabilistic planners which use static actions models.