| 研究実績の概要 |
This project was originally planned to be completed in March 2020, but was extended due through March 2021 due to the pandemic, as funds originally intended for foreign conference travel could not be spent in AY2020.
This year, we investigated best-first search using heuristic functions learned using feedforward multilayer neural networks.
We developed SING, a system which learns a neural network heuristic function for forward search-based, satisficing classical planning. SING learns distance-to-goal estimators from scratch, given a single training problem instance. Training data is generated by backward regression search or by backward search from given or guessed goal states. In domains such as the 24-puzzle where all instances share the same search space, such heuristics can also be reused across all instances in the domain. We showed that this relatively simple system can perform surprisingly well, sometimes competitive with well-known domain independent heuristics.
The neural networks can be evaluated on a GPU while the search is performed on the CPU, resulting in a heterogeneous search algorithm.
|
