Jh. Halton et Pk. Sarkar, INCREASING THE EFFICIENCY OF RADIATION SHIELDING CALCULATIONS BY USING ANTITHETIC VARIATES, Mathematics and computers in simulation, 47(2-5), 1998, pp. 309-318
This paper considers the passage of a parallel, uniform stream of part
icles through a parallel-sided, homogeneous slab of material capable o
f scattering and absorption. Particle histories are followed until the
particle escapes from the near or far surface of the slab. Each parti
cle carries a weight (initially 1, at the near surface of the slab) an
d the total weight of particles escaping through the far surface is co
mpared with the total number of particles initiated originally. To inc
rease efficiency of the computation, (i) we disregard absorption, by m
ultiplying the weight of each particle by its non-absorption probabili
ty at each collision, and (ii) we select N parallel geometric surfaces
inside the slab (parallel to the sides of the slab) and, as a particl
e crosses any such surface - if it is crossing it in a backward direct
ion - we play ''Russian roulette'' with probability l/m (multiplying i
ts weight by m if it survives), while - if it is crossing the surface
in a forward direction - we ''split'' the particle into m identical pa
rticles tall initially moving with the same velocity vector), each wit
h weight 1/m times that of the particle we split. In the case of split
ting, the paper compares two alternative techniques. Either (1) we sel
ect m independent random numbers with which to compute the m free path
lengths to the next (scattering) collision, or (2) we select one rand
om number and form m ''antithetically'' balanced values with which to
compute the m free path lengths (with appropriate weight-adjustments).
The results obtained are very promising. The use of ''antithetic spli
tting'' increases the efficiency tin terms of the work required to obt
ain a given variance in the resulting estimate of the transmission coe
fficient) by a factor ranging from about 1 to about 30 (the improvemen
t appears to be most marked for scattering probabilities around 0.5, r
ather than near 1.0). (C) 1998 IMACS/Elsevier Science B.V.