This week’s BEACON Researchers at Work blog post is by MSU graduate student José R. Llera.
My name is José R. Llera, and I received my B.S. in Computer Engineering from the University of Puerto Rico at Mayagüez. I learned about the BEACON center and their research during a visit to MSU. The study of evolutionary computation caught my interest, and I’ve been studying it at MSU ever since as a PhD student. One of the things that interest me the most about this field of study is that evolutionary computation is multi-disciplinary by nature, and you get to work with passionate people who are experts in their respective fields on an almost daily basis. This has given me the opportunity to learn exciting things from areas that would normally be outside my scope, and it opens many possibilities in solving difficult engineering problems.
One particularly exciting project was introduced to me by Dr. Goodman involving greenhouse optimization. The main motivation behind this project lies in the growing global demand for fresh vegetables, and greenhouse innovation is a hot topic for helping meet this demand. In particular, China has drafted ambitious plans to design and build a new generation of greenhouses, helping to supply its year-round needs for fresh vegetables in a way that is economically viable and environmentally friendly, as stated in its No. 1 central document of 2012, which underscores the importance of scientific and technological innovation for sustained agricultural growth.
This led to collaboration with many members from inside and outside MSU. I’m currently working directly with Dr. Goodman, Dr. Prakarn Unachak (a post-doctoral researcher at BEACON), and Chenwen Zhu (a graduate student from Tongji University, currently at MSU) in developing a system which can simulate a greenhouse environment, as well as applying evolutionary algorithms to obtain an optimized strategy for greenhouse control. An experimental greenhouse is being built in Tongji University, which is located in Shanghai, China. We expect the new models and control methodologies we are developing will be parameterized and validated at this facility. Dr. Goodman, Zhu and I visited that greenhouse in late 2012, giving us a first-hand look at its construction and operational capabilities.
A greenhouse is a complex system with interacting parameters like crops, facilities, climate and cultivation patterns, etc. How to coordinate these parameters for an efficient, productive and ecologically-safe greenhouse with a relatively optimal growing environment at the least cost has always been a research hotspot in the horticulture field. However, meeting these requirements is not trivial since in practice it’s difficult to achieve these things due to the complexity of a greenhouse environment. Our team is currently working on using a form of multi-objective evolutionary optimization (“Multi-Objective Compatible Control”, or MOCC) to allow dynamically balancing the needs for crop production vs financial and environmental costs associated with operating a greenhouse.
As for the “compatible control” part in MOCC, it is a hierarchical control strategy which takes advantage of the nature of greenhouse systems. Compromises become possible when some flexible parameters, if any, of the production system are relaxed, without damaging the whole system performance in a long term view. Such quantitative trade-offs could be made between either the economic return of the crop, the cost of maintaining and operating greenhouse facilities or the control precision of actuators in a comparatively short run.
With enough information on the greenhouse environment it’s possible with the proposed method to obtain a set of operating points, each of which is non-dominated (in the Pareto sense) by the others in the set in terms of the objectives given. That is, none of these points is better with respect to all objectives than any other point in the set. Such a set is known as a Pareto set, and is the end result of many multi-objective optimization algorithms, including MOCC. In the limit, a sequence of such Pareto sets, as more and more points are tested, is the Pareto Front, a curve along which no improvement is possible in any objective without sacrificing performance in some other objective.
Example Pareto set. Each axis represents an objective, and individuals are optimized to be as close to the origin as possible.
The best way to encapsulate, express, and implement a greenhouse system is via mechanistic models that describe the dynamics of the climate and the crop. The approach for this project uses evolutionary techniques for the optimization aspect of the problem, and given the random nature of the evolution, the final mechanistic model is must be robust enough to deal with all possible scenarios covered in the evolution process.
Typical greenhouse and variables used when modeling
Ph.D. student Bram Vanthoor of Wageningen UR Greenhouse Horticulture has developed a mathematical method for designing greenhouses that are better adapted to local conditions. We’ve found that this model is fairly comprehensive, versatile, and also tested and proven in real settings. This method has been tested for the Netherlands, Spain and USA. Although implementing all the details in Vanthoor’s model has turned out to be computationally expensive, it’s flexible enough to be adapted to our needs so that it runs with reasonable speed and accuracy.
Cucumbers and lilies are the final target crops for the experimental greenhouse. However, tomatoes are currently selected as the model crop since tomato is one of the most widely produced greenhouse vegetables in the world and knowledge about tomato yield modeling is well established. Once we have developed suitable greenhouse models and control strategies using tomatoes, it should not be too difficult to adapt them for other crops.
Currently, Dr. Unachak has been working on simplifying and speeding up the greenhouse model, and the running time for a complete growth cycle has been reduced to reasonable levels for testing evolutionary algorithms. Zhu and I are currently working on determining most appropriate aspects for optimizing our greenhouse control using NSGA-II, a type of multi-objective evolutionary algorithm. We will be able to run NSGA-II on top of the greenhouse model soon, and results that perform well could be used in the experimental greenhouse in the near future.
For more information about José’s work, you can contact him at lleraort at msu dot edu.
