Combining chemotherapy agent has been the standard treatment for most metastatic cancers. One of the experimental techniques used to find the best combination is to repetitively test the cocktails in vitro on lines of cancer cells. While this can be done on large arrays of wells in incubator, the process is time consuming and expensive. These experiments hardly investigate a very small fraction of potentially millions of cocktails. Miller and Zinner (2005) have shown that the experimental process can be drastically accelerated and improved with stochastic optimization algorithms such as Hill Climbing. The goal of this paper is two folds. First, we compare extensively several nature inspired optimization and data mining methods to find the most efficient and cost-effective drug cocktail's search. Second, we discuss the challenge of choosing the "best" fitness function, in such in vitro experiment, which is indeed not well defined. The ultimate goal of the experiment is to discover the most effective cocktail with the least complication for the cancer patients. Hence the objective function is estimated by means of domain knowledge and the knowledge gained by investigating the data obtained from preceding experiment. Our experimental study of several possible optimization techniques uses essentially benchmark problems nearby "the target application landscape" and within the experimental protocol constraints.