On run time libraries and hierarchical symbiosis

Abstract

Run time libraries (RTL) in genetic programming (GP) represent a scenario in which individuals evolved under an earlier independent evolutionary run can be potentially incorporated into a following GP run. To date, schemes for exploiting the RTL metaphor have emphasized syntactic over behavioural approaches. Thus, instructions are added to the later run such that the previous code can be explicitly indexed. In this work we demonstrate how the RTL concept is naturally supported by adopting a symbiotic framework for coevolution. Under the Pinball reinforcement learning task, we demonstrate how the initial RTL can be coevolved within a simpler formulation of the task and then used as the basis for providing solutions to a more difficult target task under the same domain. The resulting solutions are stronger than an RTL as coevolved against the target task alone or symbiosis as evolved without support for RTL.