Reflections on papers read: ORSSA 2026
Written by Robert Bennetto
At the close of the 54th ORSSA Annual Conference (27 to 30 September 2026, Devon Valley Hotel, Stellenbosch) I had the slot for "reflections on papers presented". This post is an abridged version of that talk because I can't remember exactly what I said, so I've tried to rebuild some of those comments from the notes I took on the talks.
The idea was never to review the papers or talks. It was to point out, mostly for the students in the room, where each piece of work connects to the bigger picture and what it made me think about. I'm of the opinion that habit of looking for where a problem cuts across disciplines is one of the best ones an operations researcher can develop.
This covers only the talks I happened to sit in. Sessions II (public health policy), IV (statistical methods), VII (machine learning and text analytics) and X (NLP) ran in parallel with the ones below. The absence of my commentary says nothing about the quality of the talks, just my personal preference. The full programme is still available for the curious reader.
Threads I noticed
If you take the talks together rather than one at a time, a few themes kept coming back:
The people who need us most are the hardest to model. Klara's point stuck with me most. Good decisions need good data, and the parts of society that need help most are exactly the ones AI and data-hungry methods can't reach yet. That is where OR still has to do the hard, careful work.
Humans are the difficult part of the model. Nurse dynamics (Michaela), habits and willpower (Hannah and Sean), why households adopt solar (Puja), and how schooling shapes a life (the education track). The maths is often the easier bit when compared to the soft human components.
Machine learning and classical optimisation are blending. From Talbi's co-design of hardware and algorithms, to learnheuristics for bin packing (Alice and Amahle), to PPO picking ALNS operators (Michael Tukker). The line between "ML" and "OR" is getting blurrier each year.
Fairness is becoming a first-class objective. Henri, and Jadene and Ben, both built workload equity directly into their routing models which is becoming a more pervasive theme in routing problems. It's interesting because it means that we might be at the edge of what's required to solve large routing problems so the focus is shifting to the soft components that make them actionable in the real world.
Uncertainty is front and center. Robust surgery scheduling (Michael Caie), fragile F1 strategies (Tyra), online problems where you don't know what's coming next (Alice and Amahle), and stress-testing national policy (Handre).
Research lines are long. Ideas come back years later in new hands. You rarely know at the time which conversation will change the direction of someone's work.
Day 1: Monday
Opening
Presidential address (Linke Potgieter). A lovely characterisation of operations research as a vessel for good: an altruistic force for better decisions in a world full of uncertainty. It was a good framing for many of the talks that followed.
Keynote I: El-Ghazali Talbi, Neuromorphic-based metaheuristics and computational intelligence. A nice blend of hardware and concept, pointing to architectures that are still being discovered. There is a growing emphasis on shaping the hardware to complement the algorithm. Co-design is an amazing idea. His premise was that we are at a point where we can start to define new ways to be intelligent, and maybe to improve ourselves (which is close to the definition of recursive self-improvement). I found it amusing that Benoit Mandelbrot's fractals keep turning up in the most curious places. I came away thinking that it's a wonderful and exciting time to be alive to witness the development and refinement of these technologies.
Session I: Healthcare operations
I love seeing OR applied in healthcare, and this session was a great run of it.
Michaela Licen used discrete-event simulation for nurse-centred queue management in public clinics. She did very well in the face of microphone adversity (the risk of being the first speaker of the day) and demonstrated excellent soft-skills in the way she handled herself. Distilling human dynamics into a model that can drive change is a hard and interesting problem.
Klara Engelbrecht used hybrid simulation to support patient-flow decisions in public healthcare. It brought into focus how much good decisions depend on good data. There is a part of the world that AI can't reach yet, and it's the people who need our help the most.
Michael Caie spoke on balancing efficiency and robustness in surgery scheduling. I now know to be scared of robust optimisation and the other kind of OR scheduling (operating-room scheduling).
Henri van Huyssteen showed a delightful use case for a multi-depot periodic vehicle routing problem with fairness in mobile healthcare. It's a nice example of extending and polishing the theoretical optimum into something that can be implemented.
Session III: Vehicle routing, bin packing & metaheuristics
Alice Makwangudze and Amahle Nzimande gave a wonderful breakdown of learnheuristics for online 3D bin packing. Online problems start to capture a real-world truth: you often don't have all the information at the outset when solving a prescriptive model.
Jadene Jaftha and Ben Frost tackled multi-period routing with equity, over 16 working days. An assignment problem became a routing problem, then adaptive repair. It's impressive to me that Honours students are handling routing problems of this size.
Michael Tukker used PPO-based operator selection in ALNS for a multi-depot VRP with soft time windows and heterogeneous workers. It's very interesting to watch ML and classical optimisation start to blend together like this and it highlights the skills diversity expected of OR practitioners.
Session V: Miscellaneous
Tyra Fenger modelled sequential race strategy in Formula One under uncertainty. Such an interesting environment, and a really well scoped project from start to finish (pun intended).
Mila Starling piloted module-specific AI support in OR education, and I got to learn about Mindjoy. Mindjoy has bottomless tokens, so I'm definitely going to need some of that in my life. I love the idea of giving lecturers more time back as well. Whether it improves outcomes is an open question, and a really interesting one. It was entertaining to watch the audience explore it with Mila afterwards, because so many research avenues could develop from this work.
Day 2: Tuesday
Session VI: Modelling chronic disease & health-system resources
Sean van Loggerenberg built an agent-based model of obesity, from individuals up to epidemic scale. Willpower is a fun concept to think about in the abstract. It's also interesting how limited household data forces us to hypothesise about the mechanics of social behaviour, rather than measure them directly.
Hannah Callaghan used system modelling to investigate health policies for childhood obesity. Turns out I need to form better habits as an adult to help my children form better habits growing up. Human behavioural problems are really complicated.
Isabelle Nieuwoudt presented an inventory model for mobile clinics in the Witzenberg region. A lovely problem, and I enjoy instances like this that give you a strong visual intuition for the solution: it's nice when you can "see" that it's working as expected.
Session VIII: National student competition
The standard just keeps rising. The domains are fascinating, and both supervisors and students should be proud of the work being produced at this level.
Butch Strasheim (4th year) built a decision support system for allocating solar panels and batteries. A well thought out bi-objective approach to solar decisions, and exactly the tool I wish I'd had four years ago.
Jamie Hindmarch (4th year) compared lightweight ConvNets and autoencoders for TB screening in chest X-rays. Really well thought out in terms of its practical goal (staying lightweight), with a clear understanding of its limitations.
Puja Pande (Masters) used agent-based modelling to investigate rooftop solar adoption in Cape Town. It's not just about whether the equipment is cost effective, but about the human and behavioural dynamics too. That kind of work is always harder to critique from outside, because so much intuition and subjectivity is involved.
Tinus Cloete (Masters) built predictive models for computing clusters, using Ilifu as a case study. Compute clusters are like very expensive cars, and I'm fond of the idea of tuning the engine. It's a good reminder to stay critical of systems we assume are working correctly.
Session IX: Operations research in education
Lieschen Venter's own talk, on moving from survival to academic flourishing in unequal schooling systems, set the tone. What strikes me about this track is the dedication it takes from the people on the ground to bring the work over the line. Education takes a long time to change, and a long time to measure. A wrong prescription or a misread radiology report is serious, but education shapes your trajectory and outcomes for the rest of your life. That makes it a noteworthy application of our attention.
David Coley presented his PhD plans on modelling basic education financing. He has well thought out, clear questions that will need some novel approaches to answer and defend. I'm looking forward to watching it develop.
Sha-abaan Slamang is building a prototype causal loop diagram of student belonging and success for his PhD, and is working through a lot of interesting challenges.
Keynote II: Greet vanden Berghe, Emerging challenges in personnel scheduling
I don't know enough about the details of timetabling and rostering. It was great to come away with a sense of the complexity and an appreciation of how hard these problems are. I enjoyed the recurring theme of wanting to work with other practitioners and collaborate. She also offered schools and research opportunities to ORSSA student members. My hope is that the students take up the opportunities presented to them to gain exposure in these kinds of research environments.
Day 3: Wednesday
Session XI: Optimisation for energy, infrastructure & policy systems
Solly Seeletse presented a bivariate optimisation model for value engineering in large multi-component capital projects. It speaks to another theme that keeps recurring, to me at any rate: understanding the competing, even adversarial, objectives in our environments.
Michael Bryant and Victoria Dingle used goal programming for tri-objective generator maintenance scheduling, because two competing objectives simply weren't enough. It's fun to see lines of research continue, echoing Bernie Lindner's work from many years ago. I saw them present before they finished last year, without knowing I was about to ruin them with constraint programming. That's exactly why you should present your work and lean into the unknown parts of OR: it's a vast field and we're always learning. Now we just need to convince them to look at decomposition methods to close those remaining optimality gaps.
Handre Williams is doing robust policy optimisation on a global integrated-assessment model, stress-testing South African development pathways. That is a hard problem in itself, and he did a great job of explaining how difficult these datasets and decision arenas are to work with. Handre is also an industry practitioner who keeps exploring new techniques, which is a great mindset to be in.
Robert Bennetto (me) went on a rant about all the ways he's failed this past year, in Chasing the sun: Season III. I gave a sterling example of poor time management during the presentation. Hopefully students learn from my failures on and off the pitch.
Keynote III: Luis Gouveia, Methods for the Hamiltonian p-median problem
A masterclass in exact methods and polytope analysis. Luis walked us through the "light side" and the "dark side" of the cuts for the p-median problem. Depending on how the problem data shifts, you end up with either sparse cuts or dense cuts. One family separates nicely and converges quickly. The other becomes exponentially hard to solve. The examples of delta path exchanges were wonderful, as was the use of cyclic exchanges for the cases where the simpler moves don't work. The key insight there was the relation between the delta path exchanges and the chained LKH method: working in the infeasible space rather than staying purely in the feasible space and chaining long moves together to achieve the same move operation.
I did come away wondering whether those dense cuts could be represented more compactly, perhaps in a dual form. Then again, if it were obvious, I suspect someone would have done it already! Like Greet, Luis offered opportunities to ORSSA students at his winter school in Portugal (waiving the fees entirely). I attended it many years ago, and it was a fantastic experience: a chance to interact with the best in their respective fields. I urged the students in the room to consider taking him up on his offer, because the chance to work alongside top-tier researchers like these doesn't come around often.
A note to the students
Every talk above connects to something outside its own session: routing to healthcare, behaviour to energy, scheduling to education, hardware to algorithms. Train yourself to look for those links. Present your work early, even when it isn't finished. When keynote speakers like Greet and Luis open the door to their schools and research groups, walk through it. You never know which conversation will change the direction of your work, or which idea will come back to you years from now.
Thank you to the organising committee, the session chairs, and every speaker for a great conference. Photos of the conference are available here. See you next year in the Kruger.
Comments