DEVELOPMENT OF A NEW WEIGHT WINDOW GENERATOR BASED ON MARKOV DECISION PROCESS

Abstract

A new weight window generator (WWG) is developed to get the importance estimation during the normal forward Monte Carlo (MC) simulation. This WWG is based on a reinforcement learning (RL) perspective that treat the particle transport process as a Markov decision process (MDP). The evaluation of the importance of the meshes is transformed into the evaluation of the policy value of an MDP, which comes down to solving a Bellman equation. In our tests, the transition probability matrix is tallied during the normal forward MC simulation, and then the importance is solved and converted to the WW parameters. The numerical tests demonstrate the performance of this WWG.