pgn to polyglot

[hgm]

Note that even for huge sets of games, and consequently enormous books, only a very limited set of positions near the start position would have overflowing weights. Unless the split would be selective iin terms of eco codes, the weights for these frequently visited positions should be very similar. So a quick hack where Polyglot during merge would not add the weights when one of the weights in a position that occurs in both files, but averages those when one of the weights overflows, would be completely satisfactory.

The weights are 16-bit quantities. If the split randomly samples the set of games to make the parts, the standard deviation in the weights would be the square root of thier value. So for an overflowing weight > 2^12 this would relatively to the weight be < 2^8, or 0.4%.