Building interpretable solutions for adaptive management with artificial intelligence

Jonathan Ferrer Mestres, Land and Water, Commonwealth Scientific and Industrial Research Organisation (CSIRO), Brisbane, QLD, Australia, Thomas G. Dietterich, School of Electrical Engineering and Computer Science, Oregon State University, Corvallis, OR, Olivier Buffet, INRIA, Nancy, France and Iadine Chades, EcoSciences Precinct - Dutton Park, CSIRO, Dutton Park, QLD, Australia

Background/Question/Methods
The best practice method for managing systems under uncertainty in conservation is adaptive management (AM). Most AM problems assume that the manager has complete information of the state of the system (e.g. abundance of a population) but is uncertain about its dynamics. AM problems can be modelled and solved as Mixed Observability Markov Decision Processes (MOMDPs) – a mathematical model used in Artificial Intelligence (AI). While near-optimal, a MOMDP solution is in practice difficult to understand by humans. This is because a MOMDP solution results in a large size decision graph, called a policy graph. A policy graph is akin to a decision tree where nodes represent actions to apply, and links represent observations perceived by monitoring the system states. We present the principles of our new AI algorithm that reduces the size of the policy graph to at most N nodes and K links per node. We call this new model a K-N-MOMDP where users set the values of K and N to produce easier to understand solutions while trading-off loss of performance. We evaluate our approach on two AM problems: the conservation of the endangered Gouldian finch and the control of dengue epidemics caused by Aedes albopictus mosquitoes.
Results/Conclusions
For the Gouldian finch adaptive management problem, we were able to simplify the policy graphs from 11644 nodes to 6 nodes with a loss of performance of 2%. Interestingly, the simplified policy graph allowed for the first time to derive comprehensive rules of thumb for adaptively managing the Gouldian finch. Prior to this work, these rules of thumb were incomplete and were derived through tedious simulations of scenarios. Similarly, for the Aedes albopictus adaptive control, the original policy graph had 789 nodes and up to 100 links per node. We were able to reduce the model from 100 states to K=4 states (nodes) and K=4 links per node. We retain a 91% of the original value with a reduction of 99.5% of the size of the original optimal - but undecipherable - solution.
Our approach achieved a dramatic reduction of adaptive management solutions while minimising the loss of performance. The resulting reduced K-N-policy graphs provide easier to interpret solutions for managers. In doing so, we show the value of Artificial Intelligence tools to support decision-making in ecology.