In this paper a new evaluation method for expert combination is presented. It takes into account the correlation among experts, their number and their recognition rate. An extended investigation on Majority Vote, Bayesian, Behaviour Knowledge Space and Dempster-Shafer method for abstract-level classifiers is presented. The two-way analysis of variance test and the Scheffè post-hoc comparison have been used to investigate on the factors that influence the recognition rate of the multi-expert system and to collect useful information for the multi-expert system designing.