Original paper: A general planner for agro-ecosystem models

In agro-ecosystem simulation models involving a farm, management is usually controlled through some combination of calendar- and logic-based rules. This approach has been quite successful but despite their apparent naturalness and simplicity, rules do present some difficulties in use and implementation. Even relatively simple rule sets can become quite large and difficult to follow, and they are based on the current state of the system rather than anticipating the likely future state of the system. Systems based on alternative approaches may have the capability to improve on the performance of rule-based systems and here we describe the implementation of such an alternative, called the general planner for agro-ecosystem models (GPAM), and discuss how the GPAM makes decisions in the presence of complex interactions. The GPAM works by constructing a decision tree of all possible decisions up to some defined point in the future, assessing which pathway through the decision tree leads to the best outcome, and then passing set of decisions defining that pathway to the simulation model to implement. To test its capabilities, the GPAM was instructed to control a grazing rotation in a very simple agro-ecosystem model. Tests were conducted to examine the reaction of the GPAM's performance to a range of parameters, including how far forward in time it looked when optimising the rotation length, how often it revised the rotation length plan, and how often it was allowed to change the rotation length. In all of these tests the GPAM reacted as expected and, without being provided with any prior knowledge, decided to implement the long winter rotation lengths normally used by farm managers in year-round grazing systems to make best use of limited pasture. The highest levels of animal production were obtained when the GPAM was able to make swift changes in the rotation length, when it looked further into the future when optimising the rotation length, and when it re-planned often. Initial testing indicated that the GPAM showed promise as a new way of emulating the manager in simulation models. More work is needed to assess the performance and potential of the GPAM when it is provided with imperfect or biased information on the likely future states and to allow the GPAM to manage multi-objective systems, such as when balancing production and environmental goals.