Figure 3.

(A) Diversity biased sampling and (B) fitness-biased sampling. Diversity is first maximized to generate solutions that span a given chemical space. Here, the nearest-neighbor diversity (eq 4) is defined as the average hamming distance (eq 3) to the nearest neighbor. The diverse solutions seed the property-optimizing ACSESS method, and the fitness of the diverse set is improved iteratively. The global optimum in our enumerated NKp model space was found within 30 iterations of our optimization (data shown are averaged over ten different runs).