Validating Numerical Optimization with Multiple Starts and Benchmarks

Summary

The document raises a general problem in numerical optimization: an algorithm's output may not be the global maximum or minimum, especially when the objective is difficult and the search begins from an initial vector. It frames the issue using an optimization domain represented as a unit sphere, but offers no specific objective function or finance example.

Two basic validation ideas are proposed: run the same algorithm from several starting points and compare the results, or use a different algorithm as a benchmark. These checks can reveal sensitivity to initialization or disagreement across methods, but the document does not provide a formal convergence test, guarantees of global optimality, or criteria for deciding which result is correct. Its relevance to portfolio optimization is raised as a question rather than developed into a worked method.

Key ideas

• Optimization results can depend on the algorithm's initial input.
• Repeated runs from different starting points can expose sensitivity to initialization.
• A separate algorithm can provide a benchmark for comparing candidate solutions.
• Agreement across runs does not by itself establish global optimality.

Tags

# Approaches to check/validate the output of an optimization algorithm


# Approaches to check/validate the output of an optimization algorithm












Let's say we want to optimize the a function $f(x_1,\dots, x_n)$ with $(x_1, \dots , x_n) \in \mathbb{D}^n$. For the sake of simplicity let $\mathbb{D}^n$ be the unit sphere.

We chose an optimization Algorithm $ALG_1(\vec{v})$ (here $\vec{v}$ is an input-vector necessary to intialise and run the algo. e.g. the starting-point) and apply it to our problem. Now often one does not know whether the algorithm converges to "the maximum / minimum" (on $\mathbb{D}^n$).

Thus I don't think that one would blindly accept the first output. One way of validating the result might be to run $ALG_1(\vec{v})$ with different $\vec{v}_1, \dots , \vec{v}_m$. Another could be using a different algorithm as a benchmark.

How do people in quant finance (e.g. portfolio optimization) approach this problem ?

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.