On using simulation and stochastic learning for Pattern Recognition when training data is unavailable: The case of disease outbreak

Pattern Recognition (PR) involves two phases, a Training phase and a Testing Phase. The problems associated with training a classifier when the number of training samples is small are well recorded. Typically, the matrices involved are ill-conditioned and the estimates of the probability distributions are very inaccurate, leading to a very poor classification system. In this paper, we report what we believe are the pioneering results for designing a PR system when there are absolutely no training samples. In such a scenario, we show how we can use a model of the underlying phenomenon and combine it with the principle of stochastic learning to design a very good classifier. By way of example, we consider the case of disease outbreak: Learning the Contagion Parameter in a black box model involving healthy, sick and contagious individuals. The parameter of interest involves η which is the probability with which an infected person will transmit the disease to a healthy person. Using the theory of Stochastic Point Location (SPL), the problem is reduced to a PR or classification problem in which the SPL is first subjected to a training phase, the outcome of which is used for the testing phase.