Estimate-then-optimise

Estimate-then-optimise is a two-stage decision-making framework where uncertainty is first modeled as a probability distribution using machine learning, and then integrated into a stochastic programming model for optimization. Unlike the 'predict-then-optimise' approach, which uses point forecasts, this method accounts for the full-scale uncertainty of the environment. This aligns with ISO 22301's emphasis on risk-adjusted planning and ISO 31000's principle of treating uncertainty as a core component of risk management. It is particularly relevant for industries facing high volatility, such as logistics, energy, and telecommunications, where decision-making based on average-case scenarios often leads to catastrophic failures during tail events.

Implementation typically follows three steps: 1) Data-driven distribution estimation using ML (e.g., Bayesian networks or Gaussian processes); 2) Two-stage stochastic programming to optimize the first-stage decision under uncertainty; 3) Continuous model-data feedback loops for real-time adaptation. A notable application is in airline gate assignment planning, where airlines use historical flight delay data to estimate arrival time distributions, then optimize gate assignments to minimize passenger-affecting delays. This approach has demonstrated a 20-30% reduction in delay-related costs in pilot studies. For enterprises, this translates to optimized inventory-at-risk (IaR)-based-on-demand-uncertainty, reducing stockouts by 18% and increasing service level-at-risk (SLAaR) by 12% according to industry benchmarks.

Three primary challenges exist: Data scarcity, technical complexity, and regulatory compliance. Many Taiwan SMEs lack the historical datasets required for accurate distribution estimation, making initial models unreliable. To overcome this, enterprises should adopt synthetic data generation and transfer learning techniques. Second, the need for specialized talent in both data science and operations research is high; the solution is to partner with specialized consultants like Winners Consulting Services Co., Ltd. Third, the Taiwan Personal Data Protection Act (PDPA) and GDPR impose strict limits on using operational data for ML models. Companies must implement robust data-anonymization pipelines and clear data-usage policies before deployment. The priority should be: Data-readiness assessment → Pilot implementation → Full-scale rollout within 6-12 months.

Winners Consulting Services Co., Ltd. specializes in Estimate-then-optimise for Taiwan enterprises, delivering compliant management systems within 90 days. Free consultation: https://winners.com.tw/contact