- author = "Sidney R. {Maxwell III}",
- title = "Experiments with a coroutine execution model for genetic programming",
- booktitle = "Proceedings of the 1994 IEEE World Congress on Computational Intelligence",
- year = "1994",
- pages = "413--417a",
- volume = "1",
- address = "Orlando, Florida, USA",
- month = "27-29 " # jun,
- organisation = "IEEE",
- publisher = "IEEE Press",
- keywords = "genetic algorithms, genetic programming, parallel programming, subroutines, iterative methods, coroutine execution model, synchronous parallel program execution, fitness comparison, execution time limits, iteration limits, infinite loops, infinite recursion, evolutionary progress, population tolerance",
- ISBN = "0-7803-1899-4",
- size = "6 pages",
-
URL = "
http://ieeexplore.ieee.org/iel2/1125/8059/00349915.pdf?isNumber=8059",
-
DOI = "
doi:10.1109/ICEC.1994.349915",
- abstract = "The genetic programming methodology is expanded with a coroutine model for the synchronous, parallel execution of the individual programs in the population. For certain classes of problem, namely those that support fitness comparison between individuals which are in a state of execution, this model allows the removal of execution time and iteration limits. Populations can then tolerate individuals with infinite loops (or in a suitable environment, infinite recursion), while still allowing evolutionary progress.",
-
notes = "Earlier version 11 pages available electronically. See
genetic-programming mailing list 14/12/93, 4/1/94 and
5/1/94
coroutine _model_ is described in terms of real program runtimes. Actually achieved by defining psuedo elapse time for each instruction (which is zero in some cases) and interrupting execution of the program after a certain number of these timesteps. Makes things controlable.
Run on Artificial Ant Santa Fe Trail and claims better programs produced with less effort than Koza (GP1).
Steady state pop of 1000, with 100 new individuals per cycle. Limit of 600 ticks (when comparing with \cite{koza:book}) Faster programs preferred. {"}The coroutine model found individuals which were more efficient (faster?) in solving the problem than the generational model{"} p417
Date: Mon, 24 Apr 2000 09:36:34 -0700 From: {"}Sidney R Maxwell III{"} > 1-How did Maxwell implement his method?
Basically, I executed each individual a fixed number of steps (a 'configurable' number N, with a value of as little as1). Individuals added to the population were pre-executed an appropraite number of steps to ensure that all individuals in the population had executed the same number of steps.
The problem that I was tackling was the Artificial Ant, for which evaluating fitness on partially executed individuals was meaningful.
In early experiments, I executed all individuals in the population N steps. Later, as a run-time performance enhancement, I [simply] ensured that individuals being evaluated had executed the same number of steps before comparing their fitness.
Cf. Levin search.",
