This project is only available for a summer project e.g. ASC for PhB students.
Planning and reasoning is still an unsolved problem in AI. Despite the explosion of research effort in and funding for LLMs, many works have shown lack luster results in the area of general planning and reasoning [1]. In fact, it has been shown for symbolic planning even neural networks are vastly inferior to classical algorithms or classical machine learning [2]. The reasoning for this is that neural network architectures are better suited for dealing with shallow tasks involving raw data such as image recognition and natural language interaction, whereas they struggle with long range decision making and reasoning. One can associate the two classifications of ML tasks of shallow predictions and decision making into Kahneman’s “System 1” and “System 2” halves of our brains [3].
Research Project
The work in [2] only deals with a restricted language (classical PDDL planning) for modelling decision making. More specifically, classical PDDL planning does not model stochastic actions and uncertainty. One can thus make use of a more expressive planning language such as PPDDL or RDDL. There exist work involving neural network architectures for PPDDL and RDDL (see [4] and [5], respectively) but not with classical machine techniques. Thus, we would like to extend the work of [2] to deal with more expressive planning variants given the success of classical ML over neural networks for planning. The project is rather open ended and allows for a lot of creativity.
Requirements
To undertake substantial work on this project, students should have both strong mathematical maturity and programming skills. More specifically, proficiency with working with both C++, Python and common Python ML packages (e.g. torch, scikit-learn) as well as strong results and understanding of topics in both the COMP3620/6320 (Artificial Intelligence) and COMP4670/8600 (Statistical Machine Learning) courses.
Gain
Students will be able to learn about cutting edge research in various fields of AI: machine learning (ML), automated planning, and knowledge representation (KR). The breadth and depth of the project will give students interested in pursuing a PhD a strong headstart in their academic career. ANU has a strong and welcoming planning group (both students and staff) involved in the intersection of such fields. Lastly, there is also programming experience gain from working with commonly used packages, languages and tools.
References (also suggested reading)
[1] Karthik Valmeekam, Matthew Marquez, Sarath Sreedharan, Subbarao Kambhampati: On the Planning Abilities of Large Language Models - A Critical Investigation. NeurIPS 2023
[2] Dillon Z. Chen, Felipe W. Trevizan, Sylvie Thiébaux: Return to Tradition: Learning Reliable Heuristics with Classical Machine Learning. ICAPS 2024: 68-76
[3] Hector Geffner: Model-free, Model-based, and General Intelligence. IJCAI 2018: 10-17
[4] Sam Toyer, Sylvie Thiébaux, Felipe W. Trevizan, Lexing Xie: ASNets: Deep Learning for Generalised Planning. J. Artif. Intell. Res. 68: 1-68 (2020)
[5] Vishal Sharma, Daman Arora, Mausam, Parag Singla: SymNet 3.0: Exploiting Long-Range Influences in Learning Generalized Neural Policies for Relational MDPs. UAI 2023: 1921-1931