Optimal Learning and Management of Threatened Species

Abstract

Biodiversity is being lost at an unprecedented and accelerating rate. A pressing issue is how to improve the effectiveness of conservation with limited information and resources available. We develop a Bayes-adaptive POMDP framework to optimize the learning and management of threatened species in a partially observable, dynamic, and uncertain environment. This framework takes into account uncertainties about the state, dynamics, and detectability of a species and optimizes the efforts of protection and surveying in real time. We exploit the structure of conservation problems to identify a low-dimensional hybrid belief state and reformulate the stochastic dynamic program as a piecewise deterministic optimal control problem. This enables us to obtain structural insights into the optimal policy and find an efficient approximation policy that is near optimal. Interestingly, we show that areas where the species has never been found may be more likely to contain the species than areas where it has been previously found. We derive closed-form conditions for determining whether to survey or protect a potential habitat. We conduct an empirical study on the conservation of the Hainan Gibbon and show that ignoring parameter uncertainty can significantly underestimate the presence probabilities. We further extend the model to multiple sites to optimize the spatiotemporal allocation of resources.

Keywords: Biodiversity conservation, Partially observable Markov decision processes (POMDP), Parameter uncertainty, Bayesian reinforcement Learning