Sequential Monte Carlo Methods for the Creation of Adaptive Software

In this paper we present Monte Carlo (MC) methods as a useful tool to implement adaptive behaviors. Using examples from MC integration, sequential MC methods, and importance sampling, we show the relevance of MC techniques to implement adaptive inference engines. In particular, continuing our probabilistic agent based approach presented at IWSAS 2001, we describe the particle filter as a powerful MC technique introducing important modifications to its regular operation. Our enhanced version of the particle filter includes a statistical techniques that can be used to adaptively estimate the number of particles needed by the filter without adding a significant load to its normal operation. The advantage of our approach is that the number of particles is selected at each cycle of the filter using a sound statistical criterion and thus, avoiding the use of ad-hoc thresholds. Empirical results show high accuracy of the method and also improvements over existing techniques.