Safe and Reliable Machine Learning for Planning

Machine learning can suggest good decisions in seconds, but can we trust them when lives are at stake? In this project you will build AI systems that use learned suggestions to plan flights and rescues faster, while guaranteeing that every safety and resource limit is respected. You will combine machine learning, optimisation and planning so that a learned model can be wrong without the final plan ever being unsafe.

Picture of felipe-trevizan.md Felipe Trevizan

4 Nov 2026

Background

Many decision problems must be solved efficiently and safely. A long-haul flight should use as little fuel as possible, but it must always stay close enough to an airport where it could make an emergency landing, and it must keep enough fuel in reserve. A bushfire rescue helicopter should reach evacuees quickly while avoiding smoke and fire fronts. Since weather and fire evolve unpredictably, the solution to these problems is not a fixed route but a policy: a rule saying which action to take in every situation that may arise. Constrained Stochastic Shortest Path problems (C-SSPs) are one of the main models in AI planning for these problems [1,2]. A C-SSP asks for a policy that minimises the expected value of one cost, such as flight time, while keeping other expected costs, such as fuel use, within given limits. C-SSPs can also limit the probability that something goes wrong at any point during execution, for example that the flight ever loses access to an emergency-landing site [3,4].

Constraints make planning much harder. Without them, a solver can choose the best action in each situation on its own. With them, a choice made in one situation affects whether a limit is respected overall, so the solver must reason about the policy as a whole, typically using linear programming [1,5,6]. Moreover, the best policy may need to randomise between actions (e.g. take route A 70% of the time and route B otherwise). In practice, policies that always take the same action in the same situation are preferred, but finding the best such policy is NP-hard [2,7]. At realistic scales both problems remain computationally demanding, so practitioners often fall back on conservative approximations, such as avoiding every route that could possibly become risky, which discard options that are both safe enough and much cheaper.

Machine learning can suggest good policies quickly. Models trained on small problems can be applied to much larger ones [8], and large language models can also propose policies [9]. However, nothing guarantees that their suggestions respect the constraints. A learned flight policy may favour a route whose emergency-landing site becomes unavailable under some weather conditions, and in a safety-critical setting such a policy cannot be accepted. The guiding principle of this project is that learned predictions can be useful without having to be trusted.

Aim

This project aims to turn imperfect learned policies into policies that provably satisfy all constraints of a C-SSP, faster than planning from scratch.

The student will design, implement and evaluate an approach that:

  • takes a policy suggested by a learned model (e.g. an ASNet [8], a graph neural network, or a large language model) and keeps only some of its decisions, for example by fixing the action in some situations or restricting the actions available in others, leaving the remaining decisions to the solver, similarly to how learned models help integer-programming solvers [11];
  • completes the remaining decisions with an existing C-SSP solver [1,6,10], and returns a policy only once every original constraint has been verified;
  • when completion fails or a constraint is violated, uses how much each constraint is violated and other information from the solver to decide which decisions to reopen, then repairs the policy without restarting from scratch.

The main research questions are which decisions a learned model should fix and which it should leave to the solver, whether learning suggestions that are easy to complete is more useful than imitating the choices of an existing solver, and how much inaccurate suggestions cost. Depending on the level and interests of the student, the project can focus on randomised or non-randomised policies, on limits on expected costs or on the probability of failure, or on the theory of when the returned policies are guaranteed to satisfy the constraints. The project scope can be adjusted to suit Honours, Masters or PhD level.

Ideal Student for This Project

  • computer science student with an HD average (or equivalent)
  • excited by AI research, planning, optimisation and machine learning
  • strong programming skills, ideally Python and C++
  • has taken COMP3620/6320 (Artificial Intelligence) and COMP3670/6670 (Introduction to Machine Learning) or equivalent
  • familiarity with linear programming or mixed-integer programming is a plus

Contact

If you are interested, please email Felipe Trevizan (felipe.trevizan@anu.edu.au) with:

  • the type of research project you are seeking (Honours, Masters 12-unit or 24-unit, or PhD)
  • a copy of all your transcripts
  • your CV
  • any questions you may have

References

  1. Felipe Trevizan, Sylvie Thiébaux, Pedro Santana and Brian Williams (2016): Heuristic Search in Dual Space for Constrained Stochastic Shortest Path Problems. In Proceedings of the International Conference on Automated Planning and Scheduling (ICAPS’16).
  2. Florian Geißer, Guillaume Povéda, Felipe Trevizan, Manon Bondouy, Florent Teichteil-Königsbuch and Sylvie Thiébaux (2020): Optimal and Heuristic Approaches for Constrained Flight Planning under Weather Uncertainty. In Proceedings of ICAPS’20.
  3. Peter Baumgartner, Sylvie Thiébaux and Felipe Trevizan (2018): Heuristic Search Planning With Multi-Objective Probabilistic LTL Constraints. In Proceedings of the International Conference on Principles of Knowledge Representation and Reasoning (KR’18).
  4. Ian Mallett, Sylvie Thiébaux and Felipe Trevizan (2021): Progression Heuristics for Planning with Probabilistic LTL Constraints. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI’21).
  5. Felipe Trevizan, Sylvie Thiébaux and Patrik Haslum (2017): Occupation Measure Heuristics for Probabilistic Planning. In Proceedings of ICAPS’17.
  6. Johannes Schmalz and Felipe Trevizan (2026): Efficient Constraint Generation for Stochastic Shortest Path Problems. Artificial Intelligence, 354.
  7. Johannes Schmalz and Felipe Trevizan (2024): Finding Optimal Deterministic Policies for Constrained Stochastic Shortest Path Problems. In Proceedings of the European Conference on Artificial Intelligence (ECAI’24).
  8. Sam Toyer, Sylvie Thiébaux, Felipe Trevizan and Lexing Xie (2020): ASNets: Deep Learning for Generalised Planning. Journal of Artificial Intelligence Research, 68.
  9. Subbarao Kambhampati et al. (2024): Position: LLMs Can’t Plan, But Can Help Planning in LLM-Modulo Frameworks. In Proceedings of the International Conference on Machine Learning (ICML’24).
  10. Johannes Schmalz and Felipe Trevizan (2025): Solving Constrained Stochastic Shortest Path Problems with Scalarisation. In Proceedings of ECAI’25.
  11. Vinod Nair et al. (2021): Solving Mixed Integer Programs Using Neural Networks. arXiv:2012.13349.
arrow-left bars magnifying-glass xmark