GOOSE: Learning Heuristics and Parallelising Search for Grounded and Lifted Planning
Abstract
Artificial intelligence can be categorised into two main
paradigms: model-free \emph{learners} and model-based
\emph{solvers}. Learners aim to learn functions with specified
domains and targets from data and have been popularised by major
advancements in deep learning architectures and hardware. They
are able to make quick decisions in various tasks such as
computer vision and natural language processing and are able to
handle noisy data well. However, they struggle at long range
reasoning and lack theoretical guarantees for critical tasks.
Solvers on the other hand aim to solve problems modelled by a
planning expert which require long range reasoning with
theoretical guarantees. The reasoning capabilities of solvers
come at the expense of computational complexity and difficulty of
leveraging parallelism for hardware such as GPUs. In this thesis,
we focus on a class of solvers in the form of planners, which
`plan' by finding a course of actions to taken to reach a
specified goal.
This thesis combines the best of both worlds by taking advantage
of the capabilities of learners to speedup planners for solving
large scale reasoning problems.
%
We do so by introducing our \textbf{G}raph neural networks
\textbf{O}ptimised f\textbf{O}r \textbf{S}earch
\textbf{E}valuation (\textbf{GOOSE}) framework for learning
heuristic functions for guiding search during planning.
%
The two learning tasks we focus on are learning domain-dependent
heuristic functions from small problems of a given planning
domain for use in much larger problems from the same domain, and
learning domain-independent heuristic functions, a form of
zero-shot learning where we learn heuristic functions from a set
of domains for use in problems from unseen domains.
Our contributions can be categorised into four main themes. We
\textbf{model} and construct various, novel graph representations
of both grounded and lifted planning tasks for use to learn
heuristics. The construction of such graphs are complemented with
\textbf{theory} which aim to answer the question \emph{what
domain-independent heuristics can we learn?} On the planning side
of our work, we introduce efficient \textbf{parallelisation}
techniques for speeding up heuristic search using learned
heuristic functions for planning. Our final contribution consists
of combining all our previous components into GOOSE and
evaluating it with a new and comprehensive set of experiments
which sets a new standard for the field of \textbf{learning for
planning}.