Domain-abstraction-based approach for learning multidomain planning

Abstract

Although deep learning techniques have been successfully implemented to solve domain-specific unmanned aerial vehicle planning problems, it is still a challenging task to develop a learning method to solve multidomain planning problems. Because the multidomain problems often involve learning more parameters, a dilated dataset diminishes learning speed due to its size and high dimensionality. The following two observations help tackle the issue: the state space of planning problems can be decomposed into representations of the domain state and system state, and the dimensionality problem often arises due to the huge size of the domain rather than system state. Inspired by such observations, this work presents a learning framework consisting of two networks: 1) a domain abstraction network in the form of a variational autoencoder that reduces the dimension of the domain space into a compact form, and 2) a planning network that generates a planning solution for a given domain setting. The effectiveness of the proposed learning framework is validated in case studies. © AIAA International. All rights reserved.