Decision making in a partially observable world

Do conservation managers really need to be good at counting?

How do you manage a population of sea otters that you can’t count properly? It’s a problem that goes way beyond sea otters.

Managing populations of endangered species often means making decisions and taking actions in a world of uncertainty. There are different sources of uncertainty, but one of the areas receiving attention lately is partial observability. Partial observability in a population refers to our inability to correctly count all the individuals in a population during a survey or a census. It occurs because organisms are hard to find, and it means that we have to estimate the full size of a population based on estimates from surveys. If you don’t know how many individuals there are in a population, how do you know if your management is appropriate?

These sorts of problems have been solved before using a branch of mathematics called partially observable Markov Decision processes, usually shortened to POMDP. POMDPs give optimal strategies, except they are notoriously hard to solve when the problem includes a lot of states. In the case of the sea otters, a state is the population size (number of otters). To decide what management action to take to protect the sea otters and remain realistic, you have to include a lot of states—close to one state for every possible population size. Using current methods, we don’t know how to solve that problem. It’s too big…

 

pdf-icon

FULL TEXT (PDF)

Leave a Reply