Krzysztof Krawiec


Home

Research:

edit SideBar

We consider a class of adaptive, globally-operating, semantic-based embeddings of programs into discrete multidimensional spaces termed prespaces. In the proposed formulation, the original space of programs and its prespace are bound with a learnable mapping, where the process of learning is aimed at improving the overall locality of the new representation with respect to program semantics. To learn the mapping, which is formally a permutation of program locations in the prespace, we propose two algorithms: simple greedy heuristics and an evolutionary algorithm. To guide the learning process, we use a new definition of semantic locality. In an experimental illustration concerning four symbolic regression domains, we demonstrate that an evolutionary algorithm is able to improve the embedding designed by means of greedy search, and that the learnt prespaces usually offer better search performance than the original program space.

@INPROCEEDINGS { krawiec:2011:EuroGP,
    AUTHOR = { Krzysztof Krawiec },
    TITLE = { Learnable Embeddings of Program Spaces },
    BOOKTITLE = { Proceedings of the 14th European Conference on Genetic Programming, EuroGP 2011 },
    YEAR = { 2011 },
    EDITOR = { Sara Silva and James A. Foster and Miguel Nicolau and Mario Giacobini and Penousal Machado },
    VOLUME = { 6621 },
    SERIES = { LNCS },
    PAGES = { 167--178 },
    ADDRESS = { Turin, Italy },
    MONTH = { April 27-29 },
    PUBLISHER = { Springer Verlag },
    ABSTRACT = { We consider a class of adaptive, globally-operating, semantic-based embeddings of programs into discrete multidimensional spaces termed prespaces. In the proposed formulation, the original space of programs and its prespace are bound with a learnable mapping, where the process of learning is aimed at improving the overall locality of the new representation with respect to program semantics. To learn the mapping, which is formally a permutation of program locations in the prespace, we propose two algorithms: simple greedy heuristics and an evolutionary algorithm. To guide the learning process, we use a new definition of semantic locality. In an experimental illustration concerning four symbolic regression domains, we demonstrate that an evolutionary algorithm is able to improve the embedding designed by means of greedy search, and that the learnt prespaces usually offer better search performance than the original program space. },
    KEYWORDS = { genetic algorithms, genetic programming },
    NOTES = { Part of \cite{Silva:2011:GP} EuroGP'2011 held in conjunction with EvoCOP2011 EvoBIO2011 and EvoApplications2011 },
    ORGANISATION = { EvoStar },
}


Powered by PmWiki