The first section of the paper summarizes and discusses the constraints of current evolutionary techniques used to evolve generations in current gp systems. It ends by introducing the Autoconstructive evolution technique that combines features from current evolutionary techniques in an effort to allow variation mechanisms to coevolve with the offspring to which they are applied to allowing the gp system to adapt itself to the problem space that is under consideration.
The second section introduces the Push programming language and the first autoconstructive evolution system built on Push called Pushpop. Pushpop varies from regular PushGP in that, tournaments are done comparing both parents and children to choose the bets individuals to further evolve. Pushpop also allows the system to access and execute code from other indidivduals which can result in the problem of having your solution depend on its entire population from the run in order to run.
The third section introduces AutoPush, a new autoconstructive gp system, a successor to Pushpop built on the latest iteration of the Push Programming language, Push 3. Autopush differs from Pushpop in that its aseual meaning solutions are not dependent on the entire run population. Autopush also introduces new constrants on birth and selection in an effort to make sure that children are always an improvement compared to their ancestors and that individuals constantly improving themselves have a higher chance of being selected to continue evolving.
The last sections of the paper present some results from problems solved using Autopush and concludes withe the belief that the best problem solving systems in the future will utilize autoconstructive techniques.