Regression Monte Carlo for impulse control (Q2094845)

Impulse control problems are difficult to be solved by the classical techniques using quasi-variational inequalities, because of the non-local term in the equation. The numerical algorithm in the present paper is of probabilistic-statistical type, inspired by Monte Carlo-techniques initially developed for stopping problems. It is based on the dynamic programming equation, which is discretized and solved backwards, the continuation function and the optimal impulse function being both statistically estimated through the simulation of state trajectories. Several implementations are discussed, and the algorithm is tested on two examples, treating forest rotation and irreversible investment. A package for R is publicly available on GitHub.