- author = "John R. Koza",
- title = "Genetic Programming II Videotape: The next generation",
- publisher = "MIT Press",
- year = "1994",
- address = "55 Hayward Street, Cambridge, MA, USA",
- keywords = "genetic algorithms, genetic programming",
-
URL = "
http://www.genetic-programming.org/gpvideo2.html",
-
video_url = "https://youtu.be/pRk6cth7Bpg",
-
video_url = "http://www.human-competitive.org/sites/default/files/video/genetic_programming_the_movie_part_2.mp4",
-
video_url = "http://gpbib.cs.ucl.ac.uk/koza_videos/genetic_programming_the_movie_part_2.mp4",
- size = "61 minutes 50 seconds",
- abstract = "The 1994 book \cite{koza:gp2} Genetic Programming II: Automatic Discovery of Reusable Programs extends the results of John Koza's groundbreaking work on programming computers by means of natural selection, described in this first book, Genetic Programming. This videotape provides an explanation of automatically defined functions, the hierarchical approach to problem solving by means of genetic programming with automatically defined functions, and a visualisation of computer runs for many of the problems discussed in Genetic Programming II. These problems include symbolic regression, the parity problem, the lawnmower problem, the bumblebee problem, the artificial ant, the impulse response problem, the minesweeper problem. the letter recognition problem, the transmembrane problem, and the omega loop problem.",
-
notes = "Accompanies \cite{koza:gp2}
VHS NTSC Format (KOZGV2)
VHS PAL Format (KOZGP2)
VHS SECAM Format (KOZGS2)
3:36 reproduction = copy. Crossover: two parents, two offspring. 4:20 Human DNA, massive reuse. 5:00 Lawn mower problem, move, mow, turn, jumping, 8x8 lawn. Need for solving subproblems: hierarchical reuse. Scale up. Automatically defined functions ADF. Result producing branch RPB. Dummy variable = ADF arguments. ADF1 can call ADF0. (three trees, structure preserving crossover, branch typing, STGP) 10:00 With ADFs GP evolved smaller tree solution to lawn mower: modular but not like a human solution. 12:40 comparing efficiency of approaches, eg number of fitness evaluation 14:00 E = computational effort. Efficiency ratio = 9.09 (8x8 lawn mower) structural (size) ratio = 3.65 15:40 parity functions. 3,4,5 parity with regular GP. Human odd-2-parity, 3 parity with XOR (2 parity). Terminals not used in ADF0 (only the arguments). 18:00 even 3-parity with ADF0 and ADF1. 20:00 4-parity efficiency ratio =2.18, 5 parity R_space=1.91, E=14.07, 6-parity R_space=1.77, E=52.2. 21:00 7..11-parity can be solved with ADFs. 11-parity solution using 2-parity and 4-parity (plus other stuff). 23:00 Artificial Ant on the San Mateo Trail. 23:56 solution to San Mateo Trail with ADF0. Trail divided into separate fitness cases. E=2.0. 24:30 Bumble Bee problem. Two dimensional continuous world 25 flowers. Two(?) fitness cases. With ADFs give more orderly solution. E=3.2. 25:30 Impulse response (analogue electrical circuit) controller, convolution-based fitness. E=1.46. 26:28 Obstacle avoiding robot. Aim is to mop entire 8x8 world, but avoid wasting time by bumping into posts. With ADF solution is more regular. Two(?) fitness cases. E=3.27. 27:14 Minesweeper Problem. E=6.87. 28:00 Solutions evolved by GP with ADFs very different from human solutions. 29:00 GP+ADFs smaller trees and fewer fitness cases, ie E>1. 29:09 exceptions, ie E<1. 30:40 five and six Boolean symmetry problems with and without ADFs 32:35 Structural complexity (tree size) 17 problems. 33:44 Emergent parsimony from genetic programming with ADFs. 34:02 Scaling for the Lawnmower Problems. 36:45 GP+ADFs = Scalable Automatic Programming. 36:49 Scaling for the Parity Problems. 37:59 Scaling for the Bumblebee Problems. 40:02 Wallclock Time. 40:40 Letter Recognition Problem. I v L, 6x4 pixels turtle moves, limited vision (3x3). Multiple negative cases (not I or L). 44:50 Evolved ADFs viewed as either top down or bottom up problem solution. Change of representation. 45:15 Classification in Biochemistry and Molecular Biology. Protein Bovine Pancreatic Trypsin Inhibitor BPTI. 45:42 On-The-Fly Discovery of Detectors. Swiss-prot. 46:43 alpha helix, beta strands in BPTI. 46:48 DiSulphide Bonds. 47:03 Bacteriorhodopsin transmembrane protein. 248 residues. 47:45 Transmembrane problem. 48:25 Hydrophobicity. Iteration performing branch IPB0, memory. 49:40 GP+ADF0+IPB0 slightly better then 4 human coded algorithms. 50:09 Omega Loops. 50:40 Setting up your GP, 1) terminal set, 2) function set, 3) fitness, 4) pop size, number of generations, etc., 5) when to stop, 6) hierarchy and number of ADFs. Balance excess capacity against run time resources. Retrospective analysis. 53:27 Various architectures to solve parity. 54:44 Evolutionary Method of Selecting the Architecture of the Overall Program. Point typing. 55:45 Growth of Fit Architectures. Extinction of Unfit Architectures. Emergence of a Satisfactory Architecture. 56:08 Argument Trajectory through space of possible architectures. Also 56:51 57:05 Lens Effect: Key role of representation. 57:35 Histogram of fitness of random sampling space of programs, ADFs wider spread of hits. 58:57 Summary. 60:00 Conclusion.
Koza videos \cite{koza:video} \cite{koza:video2} \cite{koza:video3} \cite{koza:video4} \cite{koza:2009:gpt}",
