- author = "Jurgen Schmidhuber",
- title = "Beyond {``}Genetic Programming{''}: Incremental Self-Improvement",
- booktitle = "Proceedings of the Workshop on Genetic Programming: From Theory to Real-World Applications",
- year = "1995",
- editor = "Justinian P. Rosca",
- pages = "42--49",
- address = "Tahoe City, California, USA",
- month = "9 " # jul,
- keywords = "genetic algorithms, genetic programming",
-
URL = "
http://www.idsia.ch/%7Ejuergen/",
-
URL = "
ftp://ftp.idsia.ch/pub/juergen/gpself.ps.gz",
-
URL = "http://citeseer.ist.psu.edu/227184.html",
- size = "8 pages",
- abstract = "Back in 1986, Dickmanns, Winklhofer, and the author used a genetic algorithm to evolve variable-length computer programs [4]. Today, our approach would be classified as {"}Genetic Programming{"} (GP). We applied it to simple tasks, including the {"}lawnmower problem{"} (later also studied by Koza, 1994). In subsequent work (1987 --- 1994), we found GP unsatisfactory for many reasons: (1) GP's way of constructing new code from old code does not improve itself: it always remains limited to the initial crossover and mutation mechanisms. (2) Like almost all other learning paradigms, GP requires concepts that are unrealistic in real world applications, such as {"}resettable environments and exactly repeatable trials{"}. In general, however, realistic environments cannot be reset -- time is one-way, and there is only one single lifelong training sequence. (3) Like almost all other learning paradigms, GP's objective function does not take into account the computation time required for learning itself. To...",
-
notes = "Presents method aiming to encourage reinforcement
learning to improve the way it learns
part of \cite{rosca:1995:ml}",
