Marketing GP

Back to main GP Brainstorming page.

GP needs to be applied to more real world problems, and when it succeeds at them should be heavily publicised. People such as Maarten Keijzer should really be blowing their own trumpets a bit more, not only to the GP community, but to the AI community at large. {Conor Ryan}

We need a 'shrink-wrapped' symbolic regression package out there ASAP. My rough outline for this is much as we discussed - free, FTPable widely, running under Unix (especially Linux, preferably using X +/- TCL), MacOS and Win95 (possibly BeBox later to be really cool). It should be limited in its function set, work mainly for straight X-Y non-linear regression, possibly 3D later, and generate TeX equations as output. It will be written in C/C++, with Unix users getting source, others executable only. Source will *not* be documented, as it is intended to be an app not a GP tutorial. Quality, robustness and good behaviour are paramount - this is to demonstrate the power and utility of GP, and must not encourage abuse leading to a bad reputation. It must be thoroughly beta-tested. It should be accompanied by a 'how-to' book (I can arrange and would like to be involved) which focuses on the app and uses that as a way of selling folk on GP as a real-world technique. {Howard Oakley}

Andy Singleton suggested that internet might be used for a globally distributed GP system based on www and JAVA. {William Langdon}

Java may provide a nice solution to two issues: 1) a standard GP platform, so researchers can try their techniques on others' problems (cf. Singleton's Problem class), and 2) distributed computation, so researchers can write their programs on one platform and try them on other, faster ones with no porting necessary -- ultimately, there could be an engine that distributes a run across many machines. {Alex Chaffee}

John Koza said: Boolean problems may be easier with typicial Machine Learning -- use fancier primitive sets to show off GP. {Eric Siegel}

List of big killer problems:

  RARS (Racing simulator)
  Electronic Circuit simulation (as in with SPICE, others)
  Financial Time Series
  RoboCup -- IJCAI - Simulated and Robotic Soccer
  Molecular Biology problems:
    protein classification
    Alpha Helix/Beta Strand/Other Classification problem
    Transmembrane Classification problems
  Image processing (satellite data, any vision processing, OCR)
  Natural Language problems
{David Andre}

Chess and other game playing programs, Natural Language Grammer Induction, Finding good primitives for vision recognition {Brij Masand}

Genuinely evolving computer virus or worm, Autonomous vehicles (robots) {Stuart Card}

I stand by chaotic time series - it remains pretty big whatever! {Howard Oakley}

Stock market prediction, voice recognition. {Robert Gukeisen}

Applications with: funding, no existing solutions, no perfect solutions yet (or theoretically impossible), NP problems. Where would NOT use GP: requirement to audit. {William Langdon}

[John Koza said] John Holland suggests the best approach for GPers is to keep on solving difficult problems. More convincing than some competition between techniques on a set of benchmark problems. {William Langdon}

Suggestion for set of GP benchmark problems. I suggested my scheduling problem http://www.cs.ucl.ac.uk/staff/wlangdon/four_node.html and said Bill Spears was doing something similar for GAs http://www.aic.nrl.navy.mil:80/~spears/functs.html {William Langdon}

We should try to find the equivalent of the NK-landscape to GP. This would allow us to measure the performance of GP regarding the complexity of the problem. {Gregory Seront}

At PPSN III, Michael Herdy and Giannino Patone used ES to solve the Rubik's Cube problem. However, they predefined operators like: two-corner-turn, three-corner-swap, ... I think this problem could be a good short-term challenge for GP. Maybe using the Hierarchical GP learning framework proposed by J. Rosca. {Hugues Juille}