What I found useful was to track when rules were used in problem solving and this showed a separation between useful rules and less useful rules.  I then removed the less useful rules knowing they could be re-learned. This produced statistically significant improvements. My publications were 2003 and 2007 on long-term learning.

Dear William George

thanks for your answer. Actually it is not the matter of omitting them, I want to merge rules. I want some methods in this area. 

Ok. Merging rules can combine either conditions or results.  Combining results would reduce the rule set, e.g., A*B->C and A*D->C becoming A*(B+D)->C, if your allow or-function as well as and conditions.  BUT, combining conditions creates a new rule and the original rules maybe less restrictive, e.g., A->C, B->C becomes A+B->C.  If you can't use or on the left, I'd have to think more about what's possible. . .