Overview
I am an Associate Professor in the School of Computing, and Director of Attention and Innovation at the Integrated AI Network.
My research is on Automated Planning, Automated Reasoning, and Machine Learning.
Student Research Projects
Artificial Intelligence Search with Application in Computer Security
- Artificial Intelligence - Satisfiability
- Artificial Intelligence - Satisfiability in HPC
- Artificial Intelligence - GPU Accelerated Satisfiability
Machine Learning with Application in Computer Security
- Learning for Combinatorial Optimisation
- Learning for Terrestrial Sky Classification with Applications in Astronomy
- Reinforcement Learning for Telescope Control
- Reinforcement Learning for Verified Tokamak Control
Co-Authored Papers and Software
Below is a summary of my collaborative research projects and co-authored publications.
- Mark Burgess (Homepage)
- With Mark Burgess and a group of our peers, we co-authored and published our distributed SAT and #SAT tool, which solves Boolean satisfiability and model counting problems in High-Performance Computing (HPC) environments.
- Project: Our tool, Dagster, is a parallel and distributed hybrid SAT solver that takes problems that are decomposed into interrelated subproblems represented by a directed acyclic graph (DAG). This structure allows for both portfolio and compositional search techniques.
- Description: The tool is demonstrated on very large combinatorial problems, including pentominoes tiling, and challenging combinatorial problems such as counting Costas arrays. We also study cybersecurity use cases motivated by protocol analysis, using CBMC to compile problem instances to SAT.
- Code (GitHub)
- AAAI Demo Paper (2023)
- PRICAI Paper (2022)
- Tutorials (YouTube)
- Ava Clifton
- My collaboration with Ava Clifton focuses on distributed and compositional implementations of Property Directed Reachability (PDR), a SAT-based search method for classical AI planning. PDR iteratively refines reachability information without needing to unroll the full system transition function. We demonstrate how to accelerate PDR in solving planning problems in parallel computing environments, and also how to exploit problem compositionality to plan using PDR in parallel.
- Project: We described the parallel-pdr planning system built using the Dagster skeleton, which implements a number of parallel computing strategies, including a novel distributed PDR algorithm designed to solve problems described in PDDL.
- Jeffrey Smith (Homepage)
- My research with Jeffrey Smith, along with instrumentation scientists Damien Gratadour and Jesse Cranney, involves Artificial Neural Network (ANN) tools for phase estimation in astronomy.
- Project: In our GANs4AO project, we apply Generative Adversarial Networks to adaptive optics. Our latest results, co-authored with Taisei Fujii, focus on Fried parameter estimation.
- Patrick Liston
- My work with Patrick Liston involves agent-based simulation studies of financial markets. We develop high-fidelity models that draw on historical limit order book data.
- Project: Our research investigates realistic agent-based market simulations that use and/or are seeded with real world tick data. We are able to study the impact of stop-loss orders and cascades, and also study familiar realistic agent categories, including pattern-based agents.
- Xiaodi Zhang
- Together with Xiaodi Zhang and Alban Grastien, I have co-authored work on Conformant Probabilistic Planning (PCP). This area focuses on finding a sequence of actions to achieve a goal with a probability above a given threshold, even with uncertainty about the initial state.
- Project: Our paper on P-CPES presents a counter-example guided PCP system. The algorithm incrementally generates plans for an increasingly refined causal abstraction until a valid plan for the PCP is found, proving particularly effective for problems requiring a high probability of success.
- Tony Allard
- Working with Tony Allard, I have focused on temporal planning, particularly on heuristics and benchmarks for problems involving durative actions.
- Project: We co-authored the TIL-Relaxed Heuristic with Patrik Haslum, which addresses the difficulty of coordinating actions in time-constrained problems. Tony implemented it in two strong temporal planners.
- Project: Tony also published the Multi-Modal Cargo Routing benchmarks.
- Mohammad Abdulaziz (Homepage)
- I have co-authored papers with Mohammad Abdulaziz, a world authority on plan-length bounds and formally verified AI planning algorithms. These papers are also with Michael Norrish. This work follows a joint paper with Jussi Rintanen.
- Project: Our work on Plan-Length Bound Computation establishes a completeness-threshold, allowing procedures like bounded model checking and planning-as-SAT to be proven complete.
- Project: We also developed the concept of a Descriptive Quotient, a method for breaking problem symmetries by solving a small, representative subproblem and generalizing the solution.
- Nathan Robinson
- My collaboration with Nathan Robinson, along with Duc Nghia Pham and Abdul Sattar, focused on advancing SAT-based planning techniques.
- Project: Our work explored two main directions in SAT-based planning. We investigated the use of Partial Weighted MaxSAT for finding cost optimal plans, and we developed a novel parallel-execution planning method using a ‘split representation’ of actions. This approach reduces the required planning horizon to find a plan, and thereby improves the efficiency of SAT-based planning.
- Other Authored Works
- In addition to the collaborations above, I have contributed decision-theoretic planning tools to plan in environments with uncertainty.
- Project: The Decision-Theoretic Planner, associated with the COGX robotics project, was designed for reasoning under uncertainty in mobile robot applications.
- Project: The Non-Markovian Reward Decision Process Planner (NMRDPP) is a system developed for planning in scenarios where rewards depend on the history of states visited, which is common in expressing complex goals. This is work with Sylvie Thiébaux.
Historical ANU TechLauncher Model of Engagement
From 2018-2024 I was convener of a cluster of professional courses at the ANU offered under the title TechLauncher. That includes the Bachelor of Software Engineering “Capstone” course at ANU. It also included students from: (i) Master of Computing, and (ii) Master of Machine Learning and Computer Vision. In this program, industry, business and government organisations engaged and collaborated with teams of early career professionals and scientists, to develop, prototype and launch real solutions.
Below are some historical artefacts from this period that folks can use to inspire development of their own practicum in Software Engineering and Artificial Intelligence.
- Learn more about the 2018/24 model via our ANU TechLauncher induction
- Get a sense of the historical ANU TechLauncher community with our showcase collateral on youtube and facebook
- ANU TechLauncher “Health Hack” innovation challenge was held at NUS, in Singapore, and we have a video package of the wrapup.
- Our 2020 ANU TechLauncher Showcase was livestreamed on the ANU Experience YouTube channel, and remains available. A fireside chat moderated by Priscilla Kan-John included invited speaker Diane Herz, at the time CEO at The Social Research Centre, and Elle McCreary (nee Syrrou), at the time Investment Director at Cardia Capital. Eleven student teams presented on topics in sustainable farming, carbon accounting, cybersecurity, and infrastructure.
Historical - ANU TechLauncher Notes
Prepared by TechLauncher facilitators and me in 2022, the below is a compilation of succinct notes regarding concepts related to the historical model of engagement.
Streamlining Project Activities
Estimation Tracking and Metrics