Suppose you have a learned policy for a Fully Observable Non-deterministic (FOND) problem. How can you be sure that it is safe? One approach is via fault analysis, which relies on algorithms that find whether individual states are safe or not. This work shows that the existing state-of-the-art algorithm for this has an exponential worst-case time complexity, and then we present a practical alternative.
ICAPS 2026 Short Paper
This short paper was accepted at ICAPS 2026!
Technical Report [arxiv]
Extended version of the paper with more experimental results in the appendix.
Source Code, Benchmarks, and Data [zenodo] [github]
The code, benchmarks, and experimental results are all here.
]]>Introducing CARL: a heuristic-search algorithm that solves Constrained Stochastic Shortest Path problems (CSSPs) optimally by solving a sequence of unconstrained Stochastic Shortest Path problems (SSPs). It fits within the framework of Lagrangian decomposition, which we frame in terms of scalarisation. What is surprising about this, is that CSSPs require stochastic policies and SSPs are solved with deterministic policies, so how can it be that we get an optimal, potentially stochastic policy from solving SSPs? The trick is that we find a particular SSP and find all its optimal deterministic policies, and these can be combined into an optimal stochastic policy for the CSSP!
ECAI 2025 Paper [link]
This paper was accepted at ECAI 2025!
Identical to the ECAI paper, but has some nice examples and pictures in the appendix.
Talk [pdf]
12 minute talk for ECAI.
Code, benchmarks, and data (coming soon)
If you need access to any of these sooner, please do get in touch!
]]>Spectral graph theory considers the matrices associated with graphs and studies these matrices’ eigenvalues and eigenvectors. Spectral graph theory has applications in many fields of computer sciences, but has had surprisingly little impact in planning. A paper by Steinerberger (2021) shows how a particular eigenvector can be used to construct a descending heuristic - following this greedily is guaranteed to lead to the goal. We give an alternative proof of this fact by showing that the eigenvector describes a network flow, and we establish further connections to planning by showing that the eigenvector describes a consistent, goal-aware heuristic. We also give some examples to illustrate the behaviour of Steinerberger’s algorithm, and answer one of his open questions.
This paper was accepted at the HSDIP 2025 workshop!
Corrigendum: the heuristic error on path graphs (fig. 5) is wrong in v1 and has been fixed in v2. Due to numeric issues, we were not selecting the smallest eigenvector but a close-to-smallest one, which does not describe a descending heuristic. The correct error curve is less complex (it’s no longer sinusoidal), but it remains unclear how to predict. Thanks /P@trik Haslum for spotting the bug!
Talk [HSDIP slides] [AmsterCAPS slides]
I gave this talk at AmsterCAPS 2025 and HSDIP 2025. The HSDIP version is newer and has more details.
Code [github]
Some python code for generating all the pictures from the paper and slides.
]]>We present a new algorithm for finding optimal deterministic policies for CSSPS called i2-dual-det and fill some technical gaps that have not been addressed before. In particular, we introduce a way to categorise how interesting a CSSPs constraints are (e.g., trivial if the constraints can be simply ignored) and we discuss how to cope with the big-M approach (it requires an upper bound that has, up to now, always been selected manually).
ECAI 2024 Paper [pdf] [link] [poster]
This paper was accepted at ECAI 2024!
Talk [slides]
I gave this talk at ANU and Saarland Uni. The slides have some examples and pictures!
Code, benchmarks, and data (coming soon)
If you need access to any of these sooner, please do get in touch!
]]>CG-iLAO* is a modification of the iLAO* algorithm, which is capable of using heuristics to ignore unpromising actions until they are needed. In our experiments, CG-iLAO* outperforms iLAO* and LRTDP (the state-of-the-art). To derive CG-iLAO* we view iLAO* under the lens of linear programming in a novel way, and generalise it with constraint generation. Then, we bring this algorithm back into the world of dynamic programming.
AI Journal 2026 Paper [pdf] [arxiv] [AIJ]
This is an extended paper that subsumes the previous ones, and fixes some minor issues.
AAAI 2024 Paper [pdf] [link] [poster]
This paper was accepted at AAAI 2024!
Technical Report [arxiv]
Extended version of the paper with more experimental results in the appendix.
1h Talk [pdf]
I gave this talk in Toulouse and at Uni Basel. It steps through iLAO* and CG-iLAO* on a toy problem, which might be helpful to gain some intuition.
Source Code, Benchmarks, and Data [zenodo] [github]
The code, benchmarks, and experimental results are all here. Note: the zenodo link points to the github repo.
]]>CoGNeRe is solver for probabilistic shortest path problems. Akin to other replanners like Robust-FF, CoGNeRe constructs a solution to the probabilistic problem by solving deterministic relaxations, which lets it return partial policies quickly in an online fashion. The novelty is that by using column generation, it can provide guarantees of optimality, and a lot of flexibility.
ICAPS DC Poster [pdf]
Poster and accompanying talks for CoGNeRe at the ICAPS 2022 Doctoral Consortium.
ICAPS DC Abstract [pdf]
4 page overview of CoGNeRe, some background, and future work. Submitted as my dissertation abstract for the ICAPS 2022 Doctoral Consortium.
Short Talk [slides]
12 minute talk that gives an overview of the algorithm and what it’s trying to do. Presented at a retreat for HDR students in the ANU Intelligence Cluster.
Honours Thesis [pdf]
100 page document that gives the background, and develops the main component of the algorithm. The formulation is a bit outdated and the experiments are problematic (the version of Robust-FF that was used for testing was hand-rolled without a lot of the optimisations in the authors’ version), but the theory is good. Submitted for the research component of my Honours Degree at ANU.
]]>