Sandra Castellanos-Paez - Learning routines for sequential decision-making

Intuitively, a system capable of exploiting its past experiences should be able to achieve better performance. One way to build on past experiences is to learn macros (i.e. routines). They can then be used to improve the performance of the solving process of new problems. In automated planning, the challenge remains on developing powerful planning techniques capable of effectively explore the search space that grows exponentially. Learning macros from previously acquired knowledge has proven to be beneficial for improving a planner's performance. This thesis contributes mainly to the field of automated planning, and it is more specifically related to learning macros for classical planning. We focused on developing a domain-independent learning framework that identifies sequences of actions (even non-adjacent) from past solution plans and selects the most useful routines (i.e. macros), based on a priori evaluation, to enhance the planning domain. 
First, we studied the possibility of using sequential pattern mining for extracting frequent sequences of actions from past solution plans, and the link between the frequency of a macro and its utility. We found out that the frequency alone may not provide a consistent selection of useful macro-actions (i.e. sequences of actions with constant objects). 
Second, we discussed the problem of learning macro-operators (i.e. sequences of actions with variable objects) by using classic pattern mining algorithms in planning. Despite the efforts, we find ourselves in a dead-end with the selection process because the pattern mining filtering structures are not adapted to planning. 
Finally, we provided a novel approach called METEOR, which ensures to find the frequent sequences of operators from a set of plans without a loss of information about their characteristics. This framework was conceived for mining macro-operators from past solution plans, and for selecting the optimal set of macro-operators that maximises the node gain. It has proven to successfully mine macro-operators of different lengths for four different benchmarks domains and thanks to the selection phase, be able to deliver a positive impact on the search time without drastically decreasing the quality of the plans.