---
title: "Bayesian optimisation | SOSX"
url: "https://sosx.ai/analysis-tools/bayesian-optimisation/"
description: "Network parameters optimised in SOSX, ranked by how much each one moves the objective, with the remaining uncertainty quantified through Monte Carlo simulation."
---

[Analysis tools](https://sosx.ai/analysis-tools/) / Testing a change

# Bayesian optimisation

“What settings give us the best outcome?”

![Bayesian optimisation in four stages: the parameter space, the guided search, the sensitivity ranking, and an optimum with its confidence interval.](https://sosx.ai/figures/analysis-bayesian-optimisation.svg)

There is a point in most programmes where the structure is settled and the argument moves to settings. How much capacity, what price, which threshold, how many of them. The search space is too large to try by hand, and the instinct to sweep every combination collapses the moment there are more than three dials.

## At a glance

What the report carries

-   The optimised parameter set
-   Sensitivity ranking
-   Monte Carlo confidence intervals

Exports to PDF or Word, with an explainability companion and the audit trail alongside it.

## The discipline

Optimisation is only as honest as the three things around it. You need an objective stated plainly enough that a number can be better or worse against it. You need to know which inputs actually matter, because optimising a parameter the objective barely responds to is a way of looking busy. And you need the uncertainty carried through, because an optimum that is only optimal at one precise set of assumptions is a liability dressed as a result.

The output worth having is not a single best answer. It is a best answer, the inputs that drove it, and the range it holds over.

## How SOSX runs it

Three methods work together. Bayesian optimisation searches the parameter space efficiently, with configurable acquisition functions so the balance between exploring and exploiting is a choice rather than a default. Sensitivity analysis ranks the parameters by how much each one moves the objective, which usually shortens the list of things worth arguing about. Monte Carlo simulation then quantifies what is left, returning confidence intervals rather than a point estimate.

It runs on the same unit-normalised parameters as every other analysis, so an objective spanning cost, time and emissions is being optimised on one lens rather than on three incompatible scales.

## What you get

The optimised parameter set, the sensitivity ranking behind it, and confidence intervals around the result. As with every analysis, the report exports to PDF or Word and carries an explainability companion documenting the conversion chains and assumptions, which matters more here than almost anywhere else: an optimisation result with no stated assumptions is a number with no argument attached.

## The others in this group

Something in the system moves, and you need to know what moves with it.

-   ### What-if analysis

    “What happens if we change this one thing?”

    A hypothetical change written into a scenario copy of the network, and the impact traced through the feedback loops.

    ![What-if analysis in four stages: the base case, one stated change, the propagation through the loops, and an impact report.](https://sosx.ai/figures/analysis-what-if-analysis.svg)

    [More information about What-if analysis](https://sosx.ai/analysis-tools/what-if-analysis/)

-   ### Deep analysis

    “How does this system actually behave over time?”

    System dynamics: the reinforcing and balancing loops, how they interact, the emergent properties no single part accounts for, and where the system is brittle.

    ![Deep analysis in four stages: stocks and flows, the loops they form, simulation over a horizon, and the behaviour that emerges.](https://sosx.ai/figures/analysis-deep-analysis.svg)

    [More information about Deep analysis](https://sosx.ai/analysis-tools/deep-analysis/)

Or see [all the analysis tools](https://sosx.ai/analysis-tools/). If you would rather we built the model and ran them for you, that is our [system research, build and analysis](https://sosx.ai/services/research-build-and-analysis/) service.

[Drop us a message and let's have a chat](https://sosx.ai/contact/)
