Learning Parameters in Non Linear Ecological Models

This paper builds on our previous work that used a non linear stochastic dynamic programming problem to solve for the optimal level of phosphorus discharged into a watershed. Typically, there is a trade off between profits from agriculture and environmental damage due to excessive levels of phosphorus (from farm animal wastes) in the watershed of a freshwater lake. The optimal management of the level of phosphorus depends not only on the costs and benefits, but also on the physical properties of the lake and on random shocks. Some lakes are more susceptible to damage from excess levels of phosphorus than others. To determine the optimal level of phosphorus for a given lake, it is necessary to learn the physical nature of the lake and its watershed. We adapted the method of optimal control with Bayes' learning of Easley and Kiefer (Econometrica 1988) to this problem. We show that in the context of the dynamic model for the flow of phosphorus in a lake, there is complete learning, i.e., there is no bandit problem in this context. We numerically solve the model and use Monte Carlo methods to solve for the mean time to learn and to determine the sensitivity of the speed of learning with respect to changes in the variance of the shocks