Parallel hierarchical evolution of string library functions

Abstract

We introduce a parallel version of hierarchical evolutionary re-combination (herc) and use it to evolve programs for ten standard string processing tasks and a postfix calculator emulation task. Each processor maintains a separate evolutionary niche, with its own ladder of competing agents and codebank of potential mates. Further enhancements include evolution of multi-cell programs and incremental learning with reshuffling of data. We find the success rate is improved by transgenic evolution, where solutions to earlier tasks are recombined to solve later tasks. Sharing of genetic material between niches seems to improve performance for the postfix task, but for some of the string processing tasks it can increase the risk of premature convergence.