Bayesian learning ecosystem dynamics with delayed dependencies from incomplete multiple source data : an application to plant epidemiology

Abstract

Ecosystem dynamics forecasting is central to major problems in ecology, society, and economy. The existing models serve as decision tools but their parameters valitity are usually not confronted to real data in a formalized approach. Dynamics bayesian network inference is promissing but limited when dealing with incomplete multiple source time series with delayed time dependencies. We propose here a temporal bayesian network with time delay and aproximate inference algorithm, to learn altogether cryptic ecosystem variables, missing data, and model parameters. The novelty in the approach is that it combines simulation-based and likelihood-based aproximate bayesian inference. The advantage of simulation based is that it allows to sample hidden processes. The advantage of likelihood based is that it provides a summary statistics that is really representing the model we are interested in. The ecosystem variables and the missing data are simulated from indicator variables using the probabilistic indicator-ecosystem model. The likelihood is estimated by averaging the probability of observed-simulated data over simulations, the parameter space is sampled with Metropolis Hasting algorithm. Another innovative proposition is to parametrize the network structure in order to learn model structure within a space provided by prior distribution. We apply to plant epidemiology.