# Identifying Effective Scenarios for Sample Average Approximation

## TL;DR

This paper proposes a novel clustering method based on Lowner-John ellipsoids to reduce sample size in Sample Average Approximation, enabling efficient distributed computation without losing solution quality.

## Contribution

It introduces a new scenario reduction technique using ellipsoid-based clustering, improving tractability of SAA with distributed computing capabilities.

## Key findings

- Reduces sample size significantly while maintaining solution quality.
- Enables distributed computation using low-cost computers.
- Achieves optimal solutions comparable to large-sample SAA.

## Abstract

We introduce a method to improve the tractability of the well-known Sample Average Approximation (SAA) without compromising important theoretical properties, such as convergence in probability and the consistency of an independent and identically distributed (iid) sample. We consider each scenario as a polyhedron of the mix of first-stage and second-stage decision variables. According to John's theorem, the Lowner-John ellipsoid of each polyhedron will be unique which means that different scenarios will have correspondingly different Lowner-John ellipsoids. By optimizing the objective function regarding both feasible regions of the polyhedron and its unique Lowner-John ellipsoid, respectively, we obtain a pair of optimal values, which would be a coordinate on a two-dimensional plane. The scenarios, whose coordinates are close enough on the plane, will be treated as one scenario; thus our method reduces the sample size of an iid sample considerably. Instead of using a large iid sample directly, we would use the cluster of low-cost computers to calculate the coordinates of a massive number of scenarios and build a representative and significantly smaller sample to feed the solver. We show that our method will obtain the optimal solution of a very large sample without compromising the solution quality. Furthermore, our method would be implementable as a distributed computational infrastructure with many but low-cost computers.