MODELS THAT LEARN TO DISTINGUISH AMONG ALTERNATIVE HYPOTHESES

For most fish populations, existing data are insufficient to distingui sh among competing hypotheses about the best harvest policy. A common approach is to seek a 'best fit' model, then manage the population as if this model were correct. Apart from the obvious risk of the 'best f it' model being incorrect, this approach fails to consider the value o f policy choices that will provide informative variation in abundance and accelerate learning about which hypothesis is correct. Modelling i s no substitute for experience, but models do help to determine what t ype of data to collect and the worth of management options. To cope wi th uncertainty about population dynamics, we advocate the use of model s that explicitly represent alternative hypotheses, the acquisition of data, and managers' response to new information. By including learnin g in simulation models, we recognize the value of experimental policie s that would help distinguish between competing hypotheses. For popula tions with discrete substocks, replicated experiments can be designed to control for environmental factors. Such management experiments have been implemented or proposed for a number of salmonid and groundfish populations.