Leading Support for Probability Management
In the February and April 2006 issues of the publication OR/MS Today, Dr. Sam Savage and coauthors Stefan Scholtes (University of Cambridge) and Daniel Zweidler (Shell Global Exploration) advance a series of ideas called Probability Management. They argue for much broader use of risk analysis methods in planning and management, especially in large enterprises, based on three ideas:
- Interactive Simulation
- Stochastic Libraries
- Certification Authority
Probability Management calls for an up-front investment in the creation of standardized Certified Distributions, created or reviewed by an expert authority, and distributed to simulation modelers, often in the form of Stochastic Libraries. The payoffs of Probability Management include:
- Much easier creation of new simulation models
- Ability to capture complex dependencies -- beyond what's possible with correlation coefficients
- 'Apples-to-apples' comparison of simulation model results
- Valid 'roll-ups' of simulation models from different groups
Risk Solver is the first product designed from the ground up to support the concepts of Probability Management. In addition to best-in-class support for Interactive Simulation, it provides direct support for Certified Distributions and Stochastic Libraries.
A Certified Distribution is a custom probability distribution, created and/or reviewed by an expert, that is made available to end user modelers as a prepackaged unit. These modelers need only the name of the Certified Distribution; they need not choose, or even be aware of, its analytic form, parameters or correlations.
An analytic Certified Distribution can be given its own random number seed, which overrides a user-specified seed for the model. This ensures that end user modelers employing the Certified Distribution will use the same samples (sequence of Monte Carlo trials) each time they run a simulation -- as long as they use the same simulation software package and version.
Given the capacity of modern computers and networks, an even better idea is to create and distribute Certified Distributions in the form of Stochastic Libraries. Such libraries contain pre-generated trial data for a group of (potentially statistically dependent) distributions. The sequence of trials is predetermined, so that if two or more end users develop and run simulation models using the same Stochastic Library, their model results can be compared and combined, on a trial-by-trial basis if necessary. The models are said to be coherent.
Risk Solver supports the use of both analytic distributions and Stochastic Libraries as Certified Distributions.