Perhaps the most straightforward way to numerically compute probabilistic I-states is to approximate probability density functions over a grid and use numerical integration to evaluate the integrals in (11.57) and (11.58).

A grid can be used to compute a discrete probability distribution that approximates the continuous probability density function. Consider, for example, using the Sukharev grid shown in Figure 5.5a, or a similar grid adapted to the state space. Consider approximating some probability density function using a finite set, . The Voronoi region surrounding each point can be considered as a ``bucket'' that holds probability mass. A probability is associated with each sample and is defined as the integral of over the Voronoi region associated with the point. In this way, the samples and their discrete probability distribution, for all approximate over . Let denote the probability distribution over , the set of grid samples at stage .

In the initial step, is computed from by numerically evaluating the integrals of over the Voronoi region of each sample. This can alternatively be estimated by drawing random samples from the density and then recording the number of samples that fall into each bucket (Voronoi region). Normalizing the counts for the buckets yields a probability distribution, . Buckets that have little or no points can be eliminated from future computations, depending on the desired accuracy. Let denote the samples for which nonzero probabilities are associated.

Now suppose that has been computed over and the task is to compute given and . A discrete approximation, , to can be computed using a grid and buckets in the manner described above. At this point the densities needed for (11.57) have been approximated by discrete distributions. In this case, (11.38) can be applied over to obtain a grid-based distribution over (again, any buckets that do not contain enough probability mass can be discarded). The resulting distribution is , and the next step is to consider . Once again, a discrete distribution can be computed; in this case, is approximated by by using the grid samples. This enables (11.58) to be replaced by the discrete counterpart (11.39), which is applied to the samples. The resulting distribution, , represents the approximate probabilistic I-state.

Steven M LaValle 2012-04-20