Figure 10

Classification results when multiple statistics are combined. The error rate decreases as more variables are used for classification. Results are averaged across branch length, number of taxa, and ratio of terminal to internal branches. Each simulation was weighted equally. Different curves correspond to different restrictions on the rate matrices used in the simulation (see Figure 6). In the "unconstrained" case, the second rate matrix Q₂ was chosen randomly, independent of the original Q₁. In the "constrained" case, each random element of Q₂ must be within 20%, 50%, or 90% of the corresponding element of Q₁. Adding variables decreased the error rate in all cases, decreasing the error rate from 50% to 27% when the rate matrix change is 20% per element and from 38% to 2.7% when the change in Q is unrestricted. The statistics added were, in order of use, statistics 37, 32, 36, and 34.