Greg Galperin noticed that rollouts can be done in parallel: You can use more than one processor to play positions out with virtually no loss of efficiency. So he’s been investigating the use of powerful parallel computers to play backgammon. This project from the 1990’s apparently came to nothing; I can’t find any references later than the ones given here.
| Jay : game learning : backgammon : rollout |
With enough processor cores, it’s possible to roll a position out in real time, creating a program that could be much stronger than current programs. This research is old, but the claim remains true in 2012.
Learning and improving backgammon strategy (1994)
Greg Galperin
Project proposal originally associated with the
Center for Biological and Computational Learning.
Proposes using rollouts to learn the evaluation function, for example, by training a neural network. It might produce better results than temporal difference learning, because the rollouts would give quite accurate evaluations.
Machine learning for prediction and control (PDF)
Greg Galperin and Paul Viola
Postscript, from
MIT AI Lab Learning & Vision Group.
From around 1997; it is not dated.
This seems to be an updated version of the above, with some results. It describes the project without giving details.
“Parallel Monte-Carlo simulation of neural network controllers”
Greg Galperin and Gerald Tesauro
Originally from MCS Division, Argonne National Laboratory.
This paper has disappeared from the Internet and seems to exist only as citations.
Proposes using rollouts to play games in real time. They were planning to create a program clearly stronger than any human.