Evo main page
GP main page
Meta-GP page
Jürgen Schmidhuber's page on the
TU Munich Cogbotlab

2nd (?) paper on Genetic Programming (1987)

(In 2010, the 25th anniversary of Genetic Programming was celebrated, and Schmidhuber gave the keynote at GP Theory and Practice 2010 @ University of Michigan's Center for the Study of Complex Systems.)

Reinforcement Learning
As an undergrad Schmidhuber used Genetic Algorithms to evolve computer programs with loops etc. on a Symbolics LISP machine at SIEMENS AG. He re-implemented the system in PROLOG at TUM. In 1987 he published this work (see jpeg scan below) together with Dirk Dickmanns and Andreas Winklhofer (authors in alphabetical order). Presumably this was world's 2nd paper on pure "Genetic Programming."

The authors were not aware of Nichael Cramer's earlier paper on pure GP (1985): A Representation for the Adaptive Generation of Simple Sequential Programs, Proc. of an Intl. Conf. on Genetic Algorithms and their Applications, Carnegie-Mellon University, July 24-26, 1985. (Stephen F. Smith proposed a related approach as part of a larger system: A Learning System Based on Genetic Adaptive Algorithms, PhD Thesis, Univ. Pittsburgh, 1980).

However, Schmidhuber was apparently the first to describe a GP implementation with loops and variable length code. Since the 1987 paper is of historic interest, one can find a jpeg scan below (source code omitted).

There may be better ways of evolving computer programs than GP's. Our contributions include Adaptive Levin Search (extending Levin's universal search algorithm, which is theoretically optimal for non-incremental search), and Probabilistic Incremental Program Evolution (PIPE). The bias-optimal way of evolving programs, however, is the Optimal Ordered Problem Solver (OOPS, J. Schmidhuber, July 2002). We are also working on learning algorithms for finding programs running on recurrent neural networks.

2011: First Superhuman Visual Pattern Recognition
First paper on Meta-GP. In the same year 1987, Schmidhuber's diploma thesis came out. Pages 7-13 are devoted to a more ambitious, self-improving, metalearning GP approach that recursively applies metalevel GP (first introduced here) to the task of finding better program- modifying programs on lower levels. The goal was to use GP for improving GP:
Page 7
Page 8
Page 9
Page 10
Page 11
Page 12
Page 13

Schmidhuber continued to work on program search and metalearning.

D. Dickmanns, J. Schmidhuber, A. Winklhofer: Der genetische Algorithmus: Eine Implementierung in Prolog. Fortgeschrittenenpraktikum, Institut f. Informatik, Lehrstuhl Prof. Radig, Tech. Univ. Munich, 1987.

Page 1
Page 2
Page 3
Page 4
Page 5
Page 6
Page 7
Page 8
Page 9
Page 10
Page 11
Page 12
Page 13
Page 14
Page 15
Page 16
Page 17
Page 18
Page 19
Page 20
Page 21

My Deep Learning since 1991