Agrizero: an O2O adversarial framework and triangulation for agricultural forecasting

2018 International Joint Conference on Materials Science and Mechanical Engineering

Chen, Yenming J.^1

Department of Logistics Management, National Kaohsiung University of Science and Technology, Kaohsiung, Taiwan^1

关键词: Adversarial networks;    Hamiltonian dynamics;    Hierarchical bayesian;    Hierarchical estimation;    Monte Carlo Markov chain;    Online auction market;    Time series clustering;    Triangulation method;   

Others  :  https://iopscience.iop.org/article/10.1088/1757-899X/383/1/012018/pdf
DOI  :  10.1088/1757-899X/383/1/012018

来源: IOP

【 摘 要 】

This study proposes an online to offline (O2O) learning framework, ArgiZero, based on three components: a generative adversarial network, an online auction market, and offline simulated agents (digital twins: buyers, farmers, natures, and markets). The generative time series and digital twins are massively generated in a manner of Monte Carlo but with extremely efficient algorithms. The goal of the generator is to produce time series that are statistically indistinguishable with the records from auction market. The goal of the discriminator is to develop a triangulation method based on semi-modeless assimilation to separate generated from actual time series. Most farmers believe agriculture is impossible to be planned because uncertainty, crowding, crisis, and risk cannot be foreseen and accommodated. This AgriZero framework alleviates the challenges through the techniques of Bayesian deep learning and data assimilation as well as the mega power of GPU computation. Thanks to Bayesian hierarchical estimation, which is akin to deep learning but more sophisticate and longer history. We are able to estimate human behaviour in the agents of buyers and farmers, natural disaster in the agents of natures, and price fluctuations in the agents of markets. The framework has been validated by a large amount of records in vegetable auctions of Taiwan and USA. The hierarchical Bayesian estimation and Monte Carlo Markov Chain particle filters used in hidden Markov model are appreciated during the massive construction of the most probable digital twins. The feature space mapping in wavelet time series, Bayesian deep learning in recurrent neural network, kernel induced Hamiltonian dynamics ABC, and hybrid SDE-kernel based forecasting for time series analysis embodies the particle generator in the GAN structure. We also apply time series clustering, RNN, bagging and boosting, and semi-modeless assimilation to assist the performance of the triangulated discriminator.