This working paper is the result of my Masters project at Imperial College London supervised by Aldo Faisal. We combine tools from grammatical inference in order to learn a grammar of actions. A Hierarchical Reinforcement Learning agent can then utilize the resulting temporally-extended actions in order to combat the curse of dimensionality in sparse reward environments.
The thesis has won the ‘Best (Applied) MAC/MRes/Specialism Project, Sponsored by Winton Capital at Imperial College London’ prize at the Department of Computing.
A subset of the full Action Grammars story (defining macro-actions from CFG production rules) also got accepted at the Cognitive Computational Neuroscience conference hosted in Berlin this year! You can find the ArXiv preprint here!