We asked 25 AI/OR researchers, tool makers, developers, and practitioners what they expect the second half of this decade will bring in the field of OR. Here is the gist of what they believe the near future holds for modeling, applications, algorithms, hardware, as well as solvers and tools.
Modeling
The biggest changes our experts expect are in the space of modeling. Their predictions can be grouped into three areas:
- The use of AI to facilitate modeling.
- The integration with other analytics and data science solutions.
- New modeling styles that will be facilitated or necessitated by novel techniques.
AI-Based Modeling
The advent of working chat bots powered by LLMs has widely conjured the vision that OR modeling may be reduced to mere textual descriptions of problems. Our experts are skeptical that this will work reliably in the near future but largely believe that textual modeling will become part of the development process.
Particularly, they point out that AI may be able to help suggest models that will solve with better efficiency and that a textual interface may help make OR solutions more adaptable, allowing for a higher level of interactivity with the user, the potential for one-of-a-kind modeling (where an optimization model is built for the sole purpose of answering one specific question), as well as real-time interaction with models and solutions.
Modeling in the Context of DS and Other Planning Software
Some experts point out that the best lever for getting OR solutions into production is to "hide" them behind already existing applications and tools in the form of existing planning software. This may help overcome both a potential lack of awareness of OR as well as skepticism that optimization will work in practice or amortize high costs for tailored solutions. Moreover, our experts expect that modeling frameworks and APIs will increasingly be tailored for the use by adjacent professions, particularly data scientists.
New Modeling Techniques
The experts believe that new optimization techniques and new requirements will drive some new trends in modeling. These regard in-model assurance, automatic explanations and documentation, solution pools for interactive causal and what-if analyses, as well as modeling geared towards new optimization heuristics, first-order methods, and new hardware (see below).
Fundamentally, the majority of our experts believe that models will become more realistic through the incorporation of distributional forecasts as well as the state of the system to be optimized, which is particularly important in sequential optimization tasks where the end-state after the execution of one plan becomes the start-state for the execution of the next.
Applications
Our experts believe that the applications that will drive OR solutions in the next five years are most likely in scheduling, vehicle routing, supply chain planning and design, capacity planning and allocation (including S&OP), and pricing. One of our experts also named robotics as a rather new field for OR that will see increased interest for optimization in the next five years.
For these applications, the experts expect that dedicated solutions will outperform general-purpose methods in the market. Particularly, they expect that solutions in these spaces will increasingly offer decision-making under uncertainty based on stochastic optimization approaches that integrate well with data science pipelines.
The rationale they give for this development is threefold:
- Data scientists will expect optimization that does not overfit predicted futures.
- Known but frequently changing scenarios require robust plans that cannot be adjusted for every scenario and therefore need to work well across the board, even when the scenarios are basically known.
- Uncertainty in system dynamics and forecasts makes more robust plans a necessity. At the same time, businesses urgently need to reduce costs and increase efficiency, creating a demand for optimization that balances expected performance and risk across the inherent uncertainty.
Algorithmic Innovation
By far the biggest impact on optimization technology is expected at the intersection of machine learning and OR. From self-improving and continuously learning optimization ("intelligent optimization") methods to LLMs that allow the interactive exploration of models and solutions to counterfactual analysis to foundation models where GNNs help learn best strategies for solving general tasks, which can then be fine-tuned for specific tasks, our experts believe that this will be the most fertile ground for new innovation in the OR space.
The second most-mentioned area of innovation is seen in the space of decision-making under uncertainty. While our experts offer little insight into the technical approaches that could fuel such innovation, they believe this will happen out of necessity, as data scientists who provide data inputs for optimization and business stakeholders are expected to require robust solutions that do not overfit point forecasts. Therefore, in contrast to ML/OR hybrids, stochastic optimization is viewed largely as a technology pull.
The third area where our experts expect development is in the area of first-order methods for optimization, particularly in order to harness the parallel compute power of GPUs. However, a major obstacle for problems with integrality constraints is considered in the current inability to effectively warm-start first-order methods after branching.
Finally, a few experts believe that there will be major strides in the explainability of optimization approaches, in inverse optimization, and in hybrid heuristic and complete optimization methods.
Hardware
Our experts closely monitor two developments in the hardware spaces: quantum computing and GPUs. While they see major obstacles that still need to be overcome before quantum computers can impact the optimization space, some believe we may see first effects already near 2030.
The impact of GPUs will be felt earlier, according to our experts. Particularly when it comes to solving pure LPs and continuous convex problems, they expect that we will soon see GPUs being used at scale.
Deployment, Tools, and Solvers
Deployment
Although some experts voice doubts, a majority believes that deployment services for optimization will become mainstream, less so because launching in the cloud was difficult but to facilitate model management, testing, and monitoring. Moreover, the experts largely believe that dedicated planning solutions with 'optimization inside' will go to market as SaaS offerings.
Solvers
On the solver side, the experts expect significant development on non-linear, heuristic, simulation-based, and stochastic solvers. Developers and practitioners voice the wish that different approaches were available as part of one modeling suite, though. The experts therefore expect that some consolidation may take place in the solver space. Also, they believe that open-source solvers will see continued development but not reach commercial solver performance.
Community
Interestingly, even though we did not introduce this topic in our prompt to the experts, many offered their view of the continued development of the OR community. The general hope is that the community will come together to advertise OR better and promote successful use cases and the positive impact that OR technology has on businesses and the environment.
Moreover, as OR technology becomes more accessible, they expect an influx of scientists from adjacent fields, particularly from data science. One expert also voiced the benefits of cross-fertilizing academia and industry, where academics provide their deep knowledge of the latest technological trends while industry experts ground the work in business reality. Finally, one expert pointed out that traditional OR methods are also used in other academic fields and that OR may gain more traction through a proliferation of OR methods in other disciplines.
To a happy and healthy 2025 in a thriving OR community!
Thanks To Our Experts
(only listing those experts from whom we received explicit permission in time before publication)
Researchers
Roberto Battiti - University of Trento
Brenda Dietrich - Cornell
Julian Hall - University of Edinburgh
Kevin Leyton-Brown - University of British Columbia
Manuel López-Ibáñez - University of Manchester
Warren Powell - Princeton
Benoit Rottembourg - INRIA
Thiago Serra - University of Iowa
Anand Subramanian - Universidade Federal da Paraíba
Kevin Tierney - University of Bielefeld
Markus Wagner - Monash
Tool Makers
Julian Hall - HiGHS
Carolyn Mooney - NextMV
Warren Powell - Optimal Dynamics
Kevin Tierney - University of Bielefeld
Petr Vilím - CoEnzyme
Developers
Julius Barth - Recentive Analytics
Claus Gwiggner - Capgemini Invent
Vijay Hanagandi - Optimal Solutions
Warren Powell - Optimal Dynamics
Joannes Vermorel - Lokad
Petr Vilím - CoEnzyme
Practitioners
Julius Barth - Recentive Analytics
Brenda Dietrich - IBM
John Fuller - Union Pacific
Claus Gwiggner - Capgemini Invent
Serdar Kadioglu - Fidelity Investments
Borja Menéndez Moreno - Trucksters
Oskar Schneider - Horvath Management Consulting
Heiko Struebing - PwC
Aristotelis Thanos - Novelis
Joannes Vermorel - Lokad
Petr Vilím - CoEnzyme