Keywords
Artificial neural networks, agent-based bioeconomic modelling, high-performance computing, sensor data, farmer decision support
Start Date
17-9-2020 3:20 PM
End Date
17-9-2020 3:40 PM
Abstract
Machine learning methods have proven to be very effective in identifying patterns and implicit dependencies in complex situations with many parameters and in providing correct classifications, predictions or decision aids with the models learned. In practice, however, the large amounts of correctly labelled training data required for such approaches are often not available. In collaboration with farm holdings in South-West Germany, we develop and test a new approach in which existing operational knowledge codified in simulation models is combined iteratively with the increasing insights of learned models. Extensive synthetic training data sets are generated by bioeconomic simulation models using high-performance computing. A learning system initiated on such data is then extended and improved by on-farm sensor data. This combination fills gaps in the existing database and enables improved training. The result is a learned, more powerful model of the observed reality with better usage potentials in crop and farm management. Our use cases consider the farm operational decisions in grain cultivation on an operational and tactical level with regard to income and environmental effects. The agent-based bioeconomic modelling system MPMAS_XN provides detailed initial simulations of the effects of fertilization and cultivation decisions both from a biological (plant growth) and an economic (expected revenue) point of view. This information is combined and compared with the results of cooperating farm holdings and with standard and average values from expert databases to fill gaps in input data and to evaluate output. Using these generated data collections, we iteratively train a suitable learning system which enables improved prediction and assessment of alternative courses of action.
Combining Machine Learning and Simulation Modelling for Better Predictions of Crop Yield and Farmer Income
Machine learning methods have proven to be very effective in identifying patterns and implicit dependencies in complex situations with many parameters and in providing correct classifications, predictions or decision aids with the models learned. In practice, however, the large amounts of correctly labelled training data required for such approaches are often not available. In collaboration with farm holdings in South-West Germany, we develop and test a new approach in which existing operational knowledge codified in simulation models is combined iteratively with the increasing insights of learned models. Extensive synthetic training data sets are generated by bioeconomic simulation models using high-performance computing. A learning system initiated on such data is then extended and improved by on-farm sensor data. This combination fills gaps in the existing database and enables improved training. The result is a learned, more powerful model of the observed reality with better usage potentials in crop and farm management. Our use cases consider the farm operational decisions in grain cultivation on an operational and tactical level with regard to income and environmental effects. The agent-based bioeconomic modelling system MPMAS_XN provides detailed initial simulations of the effects of fertilization and cultivation decisions both from a biological (plant growth) and an economic (expected revenue) point of view. This information is combined and compared with the results of cooperating farm holdings and with standard and average values from expert databases to fill gaps in input data and to evaluate output. Using these generated data collections, we iteratively train a suitable learning system which enables improved prediction and assessment of alternative courses of action.
Stream and Session
false