Evolution of Obstacle Avoidance Behaviour:Using Noise to Promote Robust Solutions

Created by W.Langdon from gp-bibliography.bib Revision:1.3872

@InCollection{kinnear:reynolds,
  author =       "Craig W. Reynolds",
  title =        "Evolution of Obstacle Avoidance Behaviour:Using Noise
                 to Promote Robust Solutions",
  institution =  "Electronic Arts",
  booktitle =    "Advances in Genetic Programming",
  publisher =    "MIT Press",
  editor =       "Kenneth E. {Kinnear, Jr.}",
  year =         "1994",
  pages =        "221--241",
  chapter =      "10",
  URL =          "http://www.amazon.co.uk/Advances-Genetic-Programming-Complex-Adaptive/dp/0262111888",
  URL =          "http://cognet.mit.edu/sites/default/files/books/9780262277181/pdfs/9780262277181_chap10.pdf",
  keywords =     "genetic algorithms, genetic programming",
  abstract =     "This chapter reports on investigations into the
                 evolution of reactive control programs for obstacle
                 avoidance in a robot like simulated vehicle for a
                 corridor-following task. The focus of these experiments
                 is the development of robust, general purpose
                 controllers; in contrast to the brittle, overly
                 specific controllers evolved in previous work. In these
                 experiments, the evolved control program is invoked
                 once per simulation step. Using proximity sensors
                 environment and steers the vehicle to avoid collision
                 with obstacles. Forward motion is constant and
                 automatic.

                 The use of noise appears to discourage brittle
                 solutions. Because those opportunistic solutions are
                 easier to evolve, discouraging them makes the problem
                 harder. In this sense, adding noise has an effect
                 similar to increasing the number of fitness trials in
                 the deterministic (non-noisy) case. It is possible to
                 over-fit a finite deterministic training set
                 Appropriate use of noise should discourage
                 over-fitting.

                 The side-effect of adding noise to the fitness test is
                 that it inevitably produces noise in the fitness value
                 determined for an individual. Fitness testing becomes
                 stochastic and repeated fitness tests of identical
                 programs yield differing results. This fitness noise
                 and the associated variance in fitness values serves to
                 mask the true fitness of an individual. When the
                 variance due to noise is comparable to the variance due
                 to genotype, the progress of evolution is markedly
                 slowed.",
  notes =        "See also \cite{SAB92:reynolds}, \cite{Alife3:reynolds}
                 but only a single vehicle rather than herd, no
                 preditor, task is to move along bendy corridor using
                 noisy sensors. {"}When the variance (in the fitness)
                 due to noise is comparable to the variance due to
                 genotype, the progress of evolution is markedly slow{"}
                 ?

                 Without noise evolved controllers brittle. THESE
                 EXPERIMENTS HAVE NOT YET PRODUCED A ROBUST CONTROLLER",
  size =         "21 pages",
}

Genetic Programming entries for Craig W Reynolds

Citations