Classical Planning in Deep Latent Space

Current domain-independent, classical planners require symbolic models of the problem domain and instance as input, resulting in a knowledge acquisition bottleneck. Meanwhile, although deep learning has achieved significant success many fields, is encoded subsymbolic representation which incompatible with systems such planners. We propose Latplan, an unsupervised architecture combining planning. Given only unlabeled set image pairs showing subset transitions allowed environment (training inputs), Latplan learns complete propositional PDDL action model environment. Later, when pair images representing initial goal states (planning inputs) given, finds plan to state latent space returns visualized execution. evaluate using image-based versions 6 planning domains: 8-puzzle, 15-Puzzle, Blocksworld, Sokoban Two variations LightsOut.