Continuous symbiotic program evolution

点击翻译按钮获取中文摘要

Genetic Programming (GP) is a subfield of Evolutionary Computation (EC) in which the general goal is to evolve programs from a population of candidate programs. Currently the applicant's research enables multiple programs from the same trial to learn to interact to establish a non-overlapping cooperative behavior during evolution. Post evolution, some subset of programs would have learnt to act under a unique subset of circumstances (a parallel framework of deployment). This project undertakes a substantial generalization in which program evolution is (1) a continuous process and (2) the interaction between programs can be hierarchical as well as parallel. Under (1) support for a continuous process of evolution assumes a competitive coevolutionary model of evolution; thus, training scenarios are coevolved with the learners (e.g., as in video game difficulty increasing as player performance improves). This also provides the basis for evolving different programs to solve different subsets of tasks, but introduces the requirement for continuous diversity maintenance in both the learner population and the population representing training scenarios. Under (2), limiting learners to a parallel form of deployment (of programs), as in ensemble style `voting' frameworks, will naturally place constraints on the types of problems against which solutions can be evolved i.e., you cannot build a new behavior from one previously learnt. Conversely, under a biological context, symbiosis has been widely acknowledged as a very important mechanism for building more complex organisms by subsuming simpler organisms in their entirety. This project pursues a symbiotic framework for hierarchical model building under GP. As additional layers of symbiosis are added, solutions evolved in earlier training scenarios are subsumed in their entirety and potentially deployed in contexts dissimilar from that in which they were originally evolved. Scalability still holds care of competitive coevolution; thus, as learners become more capable, training scenarios are sought which promote the generalization of multiple behaviours from the current population, promoting a continuous process of model building.