Publication Abstract

A recipe for Bayesian-network-driven stock assessment

T.R. Hammond

Interest in Bayesian stock assessment is rising because this simple approach is good at conveying uncertainty, at learning from experience, and at evaluating management options. Unfortunately, the algorithms driving Bayesian assessment, typically Markov Chain Monte Carlo or Sampling-Importance-Resampling, can fail to converge. While convergence is tested by various esoteric means, these can only reveal failure, never success. So, I provide a recipe for using Lauritzen-Spiegelhalter-Propagation (LSP), which drives inference in Bayesian Networks and which need not converge, instead of the typical algorithms. LSP employs discrete random variables and requires that relationships between them be specified with conditional probability tables. My recipe describes a new tool, Fuzzy Arithmetic, which permitted the use of LSP on a Schaefer stock assessment model by translating model equations into such tables. I compared LSP confidence intervals (CIs) for carrying capacity (K) to grid-search counterparts under several scenarios. LSP intervals always contained grid-search ones, though they were no more than 7% broader. Tightening the grid resolution of discrete numeric variables over regions of high posterior probability tightened LSP CIs. As LSP can handle hundreds of parameters, my recipe might be useful for almost any Bayesian problem in which slightly over-cautious CIs can be tolerated.

Reference:

T.R. Hammond (2004) A recipe for Bayesian-network-driven stock assessment. Canadian Journal of Fisheries and Aquatic Science, 61(9): 1647-1657