- author = "Hitoshi Iba and Hugo {de Garis} and Taisuke Sato",
- title = "Temporal Data Processing Using Genetic Programming",
- booktitle = "Genetic Algorithms: Proceedings of the Sixth International Conference (ICGA95)",
- year = "1995",
- editor = "Larry J. Eshelman",
- pages = "279--286",
- address = "Pittsburgh, PA, USA",
- publisher_address = "San Francisco, CA, USA",
- month = "15-19 " # jul,
- publisher = "Morgan Kaufmann",
- keywords = "genetic algorithms, genetic programming",
- ISBN = "1-55860-370-0",
-
URL = "
http://www.cs.ucl.ac.uk/staff/W.Langdon/ftp/papers/iba_1995_tdpgp.pdf",
- size = "8 pages",
-
abstract = "This paper reports an extension of STROGANOFF called
R-STROGANOFF which uses special memory terminal nodes
to provide a form of recurrancy to process time ordered
events.
All functions are polynomials (quadratics in the examples), terminals are either inputs or memories. Each memory terminals hold the value of a function node on the previous time step.
The coeffients of the polynomials are learnt by trying to match the training data using a 'Generalised Error Proporgation Algorithm'. This is determinstic. Seems like STROGANOFF's (but different?), time sequence based, based on back-propagation. The coefficients are recalculated each generation (assuming tree has changed).
Fitness function used 'minimum description length' (MDL).
Quadratic coefficients mya be limited to 0<=x<=1 to avoid divergence.
Examples: 2 step 0-1 oscilator, 4 Tomita languages (on binary alphabet).
Tree could be converted to finite state automata, which was more general than tree, ie works in all cases including those not in the training set.
On the tomita languages problems 'R-STROGANOFF works almost as well as (the best) best recurrent networks'
",
