Hill Climbing Accuracy
Starting with a couple text format books from the Gutenburg project, the files were stripped of the legal boilerplate and combined. Then a parser read the file and produced 2991 number matched files of plaintext, random key and ciphertext. Another program ran the hill climber on each file and a third generated a report on how well the hill climber performed with each ciphertext when compared with the original plaintext. I dropped this data into a spreadsheet and extracted some more data.
The average accuracy was 98.6%
greater than 70% solutions 99.93% (only 2 less than 70%)
greater than 95% solutions 93.41% (about 4-6 letters wrong)
greater than 98% solutions 76.43% (only 1 or 2 letters wrong usually JQXZ)
100% exact solutions 55.13%
Further, several scoring file systems were created and the test run again on a smaller sample. In general, the numbers had a variance of +/- about 3 or 4% and scored slightly lower which was attributed to that part of the sample being slightly more difficult or less like normal English. Otherwise, changing to log base 2 log base 1.1, log base 2 squared and even floating point log base 2 had little effect on the accuracy of the hill climber.
Limited trials with reduced randomness produced a slight speed improvement but no significant change in accuracy.
The takeaway here is that a lot of advice and general practice with cryptographic hill climbers is mostly meaningless.