Abstract ID: 134 Ant colony algorithm for driving variance reduction techniques in Monte Carlo simulations

Abstract

The use of variance reduction techniques (VRTs) in Monte Carlo calculations is mandatory in different situations in which low scoring statistics occurs, for example, when the dose deposited by radiation in small detectors in complex geometries is calculated. The efficiency of this kind of simulations could be extremely low because a huge amount of CPU time is spent on tracking particles that do not contribute to deposited dose. VRTs (e.g. Russian roulette and splitting) can improve the efficiency in this type of calculations, but unfortunately, these techniques are extremely problem-dependent, and general recipes to minimize the variance cannot be given. To solve this problem, we propose to use a method based on an ant colony algorithm that allows driving the application of VRTs with minimal user intervention and is able to increase the efficiency in low statistics Monte Carlo simulations independently of the geometry. These kind of algorithms are inspired by the behavior of actual ant colonies and permit to create optimization tools to solve complex problems such as those of finding global minima. The ant colony optimization algorithm (ACOA) here developed “learns” where to apply VRTs to favor the increase of efficiency, obtaining information from the simulation itself and generating the so-called importance maps. ACOA has been successfully employed in calculations regarding the study of dose deposition in different problems: electron beams from linear accelerators, dose deposition calculations on MOSFET detectors [1] , radiosurgery photon beams [2] , improvement of dose distribution on monoisocentric beam split technique [3] and calculations of specific absorbed fractions in non-sealed radioisotope treatments [4] . The efficiency improvements range from 10 to 100 times or even more in case of high energy photon beams.