Multi-donor Neural Transfer Learning for Genetic Programming

Genetic programming (GP), for the synthesis of brand new programs, continues to demonstrate increasingly capable results towards complex problems. A key challenge in GP is how learn from past so that successful simple programs can feed into more challenging unsolved Transfer Learning (TL) literature has yet an automated mechanism identify existing donor with high-utility genetic material problems, instead relying on human guidance. In this article we present a transfer learning which fills gap: use Turing-complete language synthesis, and neural network (NN) be used guide code fragment extraction previously solved problems injection future Using framework synthesises just 10 input-output examples, first study NN ability recognise presence fragments larger program, then end-to-end system takes only examples generates as it solves easier deploys selected solve harder ones. The NN-guided selection shows significant performance increases, average doubling percentage successfully synthesised when tested two different problem corpora, compared non-transfer-learning baseline.