# Simulation-Based Optimization
SBO uses a simulation model as the objective function for an optimizer. Instead of optimizing a mathematical formula, you optimize by repeatedly running a simulation and measuring outcomes.
The classic loop: propose new operating parameters > run simulation > measure result > propose better parameters > repeat.
Two families:
- Derivative-free methods (Bayesian optimization, evolutionary algorithms) work when you can't compute gradients. Good for continuous problems with box constraints.
- Reinforcement learning agents learn policies through trial and error in the simulated environment. Better for sequential decision-making and combinatorial problems.
The catch: derivative-free SBO on continuous problems is well-understood. Combinatorial optimization (assigning gates to flights, scheduling maintenance windows) is a fundamentally different problem class. Many companies claim "AI optimization" but are actually solving different problem types with different methods under one brand.
When evaluating claims, always ask: what type of optimization problem is this, and what method are you actually using? A company optimizing continuous process parameters and discrete scheduling problems is running two completely different systems, even if they share a UI.
Related: [[Surrogate Models]], [[Reinforcement Learning for Process Control]], [[Digital Twins]]
---
Tags: #deeptech #firstprinciple