Statistical models are not perfect, but they do help us estimate unknown complex models using fewer parameters. This paper deals with a statistical model fitted to observed actions that originate with a Markov decision process. A new Markov chain Monte Carlo (MCMC) sampler is obtained for simulation from the posterior distribution. In the authors’ own words, “this step includes a parameter expansion step, which is shown to be essential for good convergence of the MCMC sampler.” The authors then apply their method to learning a human controller in order to illustrate the concept.

One limitation of the proposed technique is that it works well, provided the data comes in batch mode and not in sequential mode, in which case sequential Monte Carlo methods are better suited.

The principle of learning to mimic a controller can be applied in diverse areas, including artificial intelligence, robotics, economics, and biology. Therefore, I recommend this paper to researchers and postgraduates in computing science and computational statistics.