From Prediction to Action: Predictive and Prescriptive AI for Complex Systems

Decision-making across healthcare, industry, finance, infrastructure, energy, transportation, and public policy increasingly relies on artificial intelligence to understand complex systems, anticipate future conditions, and support or automate action. Within this landscape, predictive and prescriptive AI represent two distinct but complementary capabilities. Predictive AI estimates future states, risks, and system behaviour, while prescriptive AI identifies effective actions through optimization, planning, control, reinforcement learning, and human-AI collaboration.

Substantial progress in both areas leaves important gaps. Predictive models often perform well on benchmark data yet struggle under distribution shift, rare events, limited data, or real operational constraints. Prescriptive models can be computationally effective but difficult to implement, insufficiently robust, or hard to validate in practice. Predictions and decisions are also frequently developed as separate systems, handed off through human judgement rather than integrated end to end. This separation limits the practical value of accurate forecasts and lets prediction errors propagate unchecked into downstream decisions. Closing that gap, through integrated predictive-prescriptive approaches, is an important and active research direction.

This Research Topic centers on integrated predictive-prescriptive approaches to decision intelligence: the use of machine learning and AI to support effective, accountable, and context-sensitive decisions in operational environments. We welcome predictive AI, prescriptive AI, and integrated predictive-prescriptive research, with the strongest fit for contributions that connect the two. Predictive contributions should demonstrate relevance to downstream decisions, not accuracy improvements alone, and may address forecasting, risk estimation, anomaly and failure detection, behavioural modelling, system-state estimation, predictive maintenance, and uncertainty quantification. Prescriptive contributions should engage with uncertainty, dynamic information, or data-driven decision processes, and may address optimization under uncertainty, reinforcement learning, simulation-based optimization, planning, scheduling, routing, and control. Integrated contributions may examine how predictions inform decisions, how operational objectives shape learning, how uncertainty propagates from forecasting to action, or how feedback supports adaptation over time. We encourage the use of decision-oriented metrics, such as utility, cost, regret, robustness, latency, and operational feasibility, alongside standard predictive or optimization measures. References to large language models and agentic AI are welcome where they support decision-making within the Topic’s core focus on integrated predictive-prescriptive methods.

We particularly encourage studies grounded in real-world systems that address imperfect data, latency, distribution shift, institutional constraints, trust, adoption, conflicting objectives, feedback effects, and human behaviour. The Research Topic is open across healthcare, energy, transportation, logistics, manufacturing, agriculture, finance, infrastructure, environmental systems, and public services.

Relevant topics include:
- Integrated predictive-prescriptive approaches and decision-focused learning
- Predict-then-optimize and end-to-end decision pipelines
- Uncertainty propagation from prediction to action
- Forecasting, risk estimation, and uncertainty quantification, with demonstrated relevance to downstream decisions
- Predictive maintenance and anomaly detection
- Robust, explainable, and transferable predictive AI
- Combinatorial optimization, planning, scheduling, routing, and resource allocation under uncertainty
- Constraint-aware learning and constraint-based decision-making
-Reinforcement learning and multi-agent decision-making
- Simulation-based optimization and digital twins
- Program synthesis and probabilistic logic programming for planning and decision systems
- Decision-oriented evaluation: utility, cost, regret, robustness, latency, and operational feasibility
- Human-in-the-loop AI and trustworthy AI
- Deployment studies and operational evaluation
- Large language models and agentic AI in support of decision-making