Dynamic programming and learning models for management of a nonnative species

Nonnative invasive species result in sizeable economic damages and control costs. Because dynamic optimization models break down if controls depend in complex ways on past controls, nonuniform or scale-dependent spatial attributes, etc., decision-support systems that allow learning may be preferred. We compare two models of an invasive weed in California's grazing lands: (i) a stochastic dynamic programming model and (ii) a reinforcement-based, experience-weighted attraction (EWA) learning model. We extend the EWA approach by including stochastic forage growth and penalties for repeated application of environmentally harmful controls. Results indicate that EWA learning models offer some promise for managing invasive species.