An alternative would be to try out various illegal moves/switches and then simulate set of games from that starting point to see whether it is still roughly even. For example, I assume pre-computers many people had tried playing with the pawns on the last row and the other pieces on the second row.

]]>All this is reminiscent of the Markov Chain Monte Carlo literature on efficient sampling, the goal here is to sample from the "score=" region of a VERY high dimensional parameter space. If you generate PURE random moves, even if you disallow capture, you'll probably be unlikely to be in a "neutral" position.

]]>Perhaps this variant should be called (Multivariate Chess Markov Chain Monte Carlo) That's MC-MCMC for you acronym buffs.

]]>