A New Experimental Analysis

Our third contribution is an experimental analysis of the factors that affect the performance of NMRDP solution methods. Using NMRDPP, we compared their behaviours under the influence of parameters such as the structure and degree of uncertainty in the dynamics, the type of rewards and the syntax used to described them, reachability of the conditions tracked, and relevance of rewards to the optimal policy. We were able to identify a number of general trends in the behaviours of the methods and provide advice concerning which are best suited in certain circumstances. Our experiments also lead us to rule out one of the methods as systematically underperforming, and to identify issues with the claim of minimality made by one of the PLTL approaches.

Sylvie Thiebaux 2006-01-20