Evolutionary Informatics - Study of Information Processing in Evolutionary Systems

Study of Information Processing in Evolutionary Systems

The notion that information processing is essential to life and to evolution predates the entry of the term informatics into the English language (1966). Various investigators argued in the 1940s that certain principles of information processing apply both in living and engineered systems, and much of their thinking is encapsulated in Norbert Wiener's Cybernetics, or Control and Communication in the Animal and the Machine (1948). Wiener regarded evolution as phylogenetic learning, or accrual of information in the genome. While cybernetics and biocybernetics address information, they place an emphasis on principles of feedback and control that informatics does not. Relatively recent work has focused on evolution as optimization of fitness functions, and has addressed the role of information in optimization. Beginning with a 1995 technical report and continuing with a 1997 article, "No Free Lunch Theorems for Optimization" Wolpert and Macready established that evolutionary algorithms have average performance no better than that of random search. They argued that superior performance could be achieved only if algorithms incorporate prior knowledge of problems, and provided an information-geometric analysis of how algorithms and problems are matched (and mismatched). English argued in 1996 that there was no free lunch due to an underlying "conservation of information," and pursued the notion further in 1999. In that work, conservation was characterized in terms of Shannon information and mutual information. In 2000, English turned to Kolmogorov complexity as a measure of information in instances of fitness functions and optimization algorithms. He observed that almost all problems exhibit a high degree of Kolmogorov randomness, and thus are easy for almost all optimization algorithms. In 2004, English gave a new perspective on conservation by way of characterizing approximate satisfaction of a necessary and sufficient condition for "no free lunch." Wolpert and Macready proved the existence of coevolutionary "free lunches" in 2005. This may be interpreted as the discovery of a problem class for which some coevolutionary algorithms are generally better informed than others of how to solve problems.