Probabilistic reasoning and multiple-expert methodology for correlated objective data

In this paper, a numerical expert system using probabilistic reasoning with influence structure generated from the observed data is demonstrated. Instead of using an expert to encode the influence diagram, the system has the capability to construct it from the objective data. In cases where data are correlated, instead of compromising the performance by wrestling with different influence structures based on the assumption that all the environment variables are observed, we incorporated the flexibility of including unobservable variables in our system. The resulting methodology minimised the intervention of a domain expert during modelling and improved the system performance. Global optimisation using all variables is often very difficult and unmanageable in probabilistic network construction. In our approach, we group all the variables into subsets and generate advice for these subsets of features using multiple small probabilistic networks, and then seek to aggregate these into a consensus output. We proposed a probabilistic aggregation using the joint probability of data and model approaches. In this approach, we avoided the very high-dimensional integration over all possible parameter configurations. The resulting system has the benefit of a multiple-expert system and is easily expandable when new information is to be added.