An adaptive architecture for portability of greenhouse models

Abstract

This work deals with the portability of greenhouse models, as we believe that this is a challenge to their practical usage in control strategies under production conditions. We address this task by means of adaptive neural networks, which re-adjust their weights when transferred to new conditions. Such an adaptive account for computational models is typical of the field of developmental robotics, which investigates learning of motor control in artificial systems inspired on infants development. Similarly to robots, greenhouses are complex systems comprising technical and biological elements, whose state can be measured and modified through control actions. We present an adaptive model architecture to perform online learning on greenhouse models. This learning process makes use of an episodic memory and of online re-training. This allows for adaptation without the need for a complete new training, which might be prohibitive if the data under the new conditions is scarce. Current experiments focus on how a model of tomato photosynthesis, developed in a research facility, can adapt itself to a new environment in a production greenhouse. Further research will focus on model plasticity by means of adaptive learning rates and management of the episodic memory described in this paper. The models presented as a proof-of-concept estimate the transpiration and photosynthesis of a hydroponic tomato crop by using measurements of the climate as inputs. The models are trained and tested using data from a greenhouse in Berlin, Germany. Thereafter, the adaptive architecture is fed with data from a production greenhouse in southern Germany, where other tomato varieties were grown under different irrigation and climate strategies. The proposed adaptive architecture represents a promising tool for spreading the use of models produced by high-tech research centers to the greenhouse production sector.