Evolving and Generalising Very High Quality Control Knowledge for AI Planning

This project is concerned with making AI planning practical by exploiting evolutionary learning techniques to acquire control knowledge automatically. This control knowledge can be used to prune useless branches of the search space and to propose promising branches. Whilst human-coded control rules have been shown to be useful in planning, their specification is an effort-intensive process which makes them difficult or impossible to generalise. Their use has tended to be confined to relatively simple domain models about which the human has a good understanding of the dynamics. We propose to learn powerful rules automatically from both static and dynamic sources of information about the planning domain and the process of planning within that domain. Furthermore, we will develop a method for learning generic rules that apply to classes of domains and that can be automatically customised for a particular domain. This reduces or even removes the burden on the human and results in scalable planning technology.