Online Learning For AI Planning (OLP)
Until the application portal opens, please send questions to Tomáš Pevný and Jendrik Seipp
See the CTU job portal for information about requirements (for a previous iteration of the call).
The symbolic search in large state spaces is one of the foundational problems in AI. Modern solvers typically use some variant of graph search algorithm with a strong heuristic function. Not surprisingly, the heuristic function is nowadays frequently realized by neural networks. The problem is that the training set usually only contains a few small problem instances, while the testing set contains large instances, which inevitably leads to issues with generalization. This project will therefore investigate options how to adapt the heuristic function during the search to efficiently use information from states visited during the search. We anticipate two issues: computational complexity and construction of the feedback signal from the explored state space. The ultimate goal is to make the resulting planning system outperform planners that use hand-coded heuristics.
PI: Tomáš Pevný
Core team: Jendrik Seipp + one postdoc
Funding: This project is supported by the CROP Postdoctoral Fellowship Programme at the Czech Technical University in Prague.