Selecting Descriptors by Evolution

It has been reported that models developed with selected subsets of descriptors can be more accurate and robust that those using all possible descriptors [84]. To select the appropriate descriptors for each of the four classes of protons, we can compare the prediction results for a selected set of descriptors to those obtained when all the descriptors are used. GAs are an excellent tool for this task. The general approach of the GA has been briefly introduced. Detailed explanations can be found in the original publications, and a comprehensive review on GAs is available in the literature [85]. We will focus here on applying GAs in the context of NMR shift prediction. 

FIGURE 6.19 Specification of a subset of descriptors from a pool of possible descriptors by a chromosome. A cross-over mutation leads to an exchange of descriptor availability in two chromosomes. 

The GA is now simulating an evolution with the entire population. The size of the population corresponds to the number of selected descriptor subsets. Each population undergoes now an evolution for a fixed number of generations  

This process continues until a termination criterion is reached either a maximum number of generations or a satisfactory fitness level for the population. 

---