Deep symbolic learning and planning
Project Leader: Adrian Pearce
Primary Contact: Adrian Pearce (firstname.lastname@example.org)
Keywords: artificial intelligence; autonomous systems; machine learning
Disciplines: Computing and Information Systems
At the heart of this PhD challenge is linking symbolic AI planning to deep learning techniques. This PhD project tackles the fast emerging endeavour of integrating symbolic AI planning with data-driven deep learning, which have recently been shown to be nice complementary approaches. It will build on AI planning techniques developed at the University of Melbourne in automated planning – demonstrated to be competitive with DeepMind for playing Atari in the Arcade Learning Environment (ALE) – and expertise in deep machine and statistical learning and natural language processing. Specifically, the thesis will involve linking a symbolic processor with a neural network, to meet the competing demands of the amount of data available for deep learning techniques, with the need for symbolic interpretation. Relevant applications include AI dialogue systems which must integrate a natural language interface to a logical processor for disambiguation queries, e.g. ready for an SQL server. Recent model-free planning techniques can be harnessed, where factored state representations of a simulator can be used to capture more complex domain knowledge, while facilitating a step-change in reinforcement learning performance. The thesis will utilise learning for recognising context and ranking policies (similar to AlphaGo) while utilising AI planning for symbolic query interpretation. You will be involved in developing a new approach, harnessing reinforcement learning and self-learning, using simulators and query servers.