Abstract: | This dissertation describes the adaptation of credibility theory, an actuarial technique, to the domain of intelligent agents. Credibility theory is used for experience rating in general insurance; here it is adapted for experience rating of an agent's estimate of the dynamism of its environment. The goal is to minimise the cost/benefit of making observations by relating the observation frequency to the dynamism of the world.
The results from adapted credibility theory are compared to those achieved by a simple implementation of a Kalman filter, but the testbed does not provide conclusive results. However, the adaptation of a credibility theory is shown to be flawed, since variable weighting of the a priori data produces a performance deterioration compared to a fixed weighting between a priori and recent estimates.
Various problems with the adaptation of credibility theory have been identified and further work suggested to improve the performance.
|