[jasone]

This was the idea that I kicked around for years before eventually writing a Scrabble engine in 2018. But by the time I got to the point where Monte Carlo simulation of the opponent's rack leave became feasible, I questioned the utility. Part of the problem is that if your opponent knows you incorporate such a simulation into the analysis, it's possible to subvert the analysis. And the most obvious counter is to not do the simulation. Another issue is that the advantage gained by opponent rack leave simulation is subtle, and it's costly to run enough games to measure the overall utility. I ended up instead putting my effort into controlling search depth termination with uncertainty estimation (spend more time on the portions of the tree that are more volatile).