Figure 3 shows the optimal value function approximation
obtained by HALP and AVI and refers to the case when we
use one Gaussian function inside Ax and one outside Ax, for
each mode. The 2-dimensional Gaussian functions consist
of the product of two univariate Gaussian functions whose
means are centered in the middle of [xl
, xu], while standard
deviations are set equal to 1 for the univariate Gaussian
functions with support in Ax and to 15 for those with support
outside Ax.
Despite the simple approximation adopted in the HALP
method for the value function, when the initial state s(0) is
uniformly distributed over A, the estimates of the probability
of remaining within the set A for the whole time horizon
[0,N] are quite close (0.9526 for the HALP solution and
0.96 for the AVI solution).
Table II reports the time needed by CPLEX for Matlab
running on a Intel Xeon E5-2630 with 64 GB of RAM
to determine the HALP approximation to the optimal value
function once the integrals involved in the LP program (11)
have been computed