Elsevier

Parallel Computing

Volume 27, Issue 5, April 2001, Pages 571-585
Parallel Computing

Evolving two-dimensional cellular automata to perform density classification: A report on work in progress

https://doi.org/10.1016/S0167-8191(00)00078-8Get rights and content

Abstract

We present results from experiments in which a genetic algorithm (GA) is used to evolve 2D cellular automata (CA) to perform a particular computational task (“density classification”) that requires globally coordinated information processing. The results are similar to that of earlier work on evolving 1D CAs. The behavior of the evolved 2D CAs is analyzed, and their performance is compared with that of several hand-designed 2D CAs.

Introduction

In many natural systems, simple, locally interacting components give rise to coordinated global information processing. In both natural and human-constructed information-processing systems, allowing global coordination to emerge from a decentralized collection of simple components has important potential advantages – e.g., speed, robustness, and evolvability – as compared with explicit central control. However, it is difficult to design a collection of individual components and their interactions in a way that will give rise to useful global information processing or “emergent computation”. The term “emergent computation” refers to the appearance in a system's temporal behavior of information-processing capabilities that are neither explicitly represented in the system's elementary components.

In order to understand the mechanisms by which an evolutionary process can discover methods of emergent computation, a simplified framework was proposed and studied by Crutchfield, Mitchell, and their colleagues [2], [3], [8] in which a genetic algorithm (GA) evolved 1D cellular automata (CA) to perform computations. In their work the GA was able to discover CAs with high performance on tasks requiring cooperative collective behavior. In this paper we describe extending this work to 2D CAs.

Section snippets

A density-classification task for cellular automata

In [8], 1D CAs were evolved by a GA to perform a density classification task called the “ρc=1/2” task. (This built on earlier work by Packard [9]). A successful CA for this task will decide whether or not the initial configuration (IC) contains more than half 1s. More formally, let ρ0 denote the density of 1s in the IC. If ρ0>0.5 then within M time steps the CA should go to the fixed-point configuration of all 1s; otherwise within M time steps the CA should go to the fixed-point configuration

Details of experiments

We used a GA to evolve 2D, binary state CAs to perform the ρc=1/2 task. GAs are search methods inspired by biological evolution. In a typical GA, candidate solutions to a given problem are encoded as bit strings (“chromosomes”). A population of such strings is chosen at random and evolves over several generations under selection, crossover and mutation. At each generation, the fitness of each bit string is calculated according to some externally imposed fitness function, and the highest-fitness

Results

We performed more than 100 different runs of the GA with the following parameters: for each CA in the population P=100; E=10; I=100; m=0.016; G=100 (in some runs G was set to 400), each with a different random-number seed. The dynamics of a typical run is shown in Fig. 6, which plots the best fitness rule, the elite mean fitness and the population mean fitness versus the generation for the first 50 generations. Before the GA discovers high fitness rules, the fitness of the best CA rule

Conclusion

To date, the best 2D CA rules evolved by the GA in our experiments implement unsophisticated block-expanding strategies. Although they have higher performance than that of simple hand-designed rules such as the majority or the anneal-rule, their performance is lower than that of the 2D GKL rule and the non-local majority rule, and that of the high-performance rules evolved by the GA in the 1D case [2], [3]. Why did the GA not find higher-performance rules with much better performance? Some of

Acknowledgements

Many thanks to Rajarshi Das and Wim Hordijk for their assistance and suggestions. This research was supported by the Santa Fe Institute, under grants NSF-IRI-9705830 and ONR N00014-95-1-1000. FJM was also supported by the Spanish Ministerio of Educacion y Ciencia project no. PB97-0741.

References (10)

There are more references available in the full text version of this article.

Cited by (43)

  • Evaluation of dynamic behavior forecasting parameters in the process of transition rule induction of unidimensional cellular automata

    2010, BioSystems
    Citation Excerpt :

    However, the opposite does not hold, since a many rules that present a “null” behavior are not solution for the classification by the majority problem. Several authors have proposed the use of computational intelligence techniques (more specifically, evolutionary computation), for determining appropriate transition rules for the classification by the majority rule problem (Richards et al., 1990; David et al., 1996; Juillé et al., 1998; Morales et al., 2001; Ferreira, 2002; Oliveira et al., 2002b). The general approach used by those works is the evolution of a population of individuals (in this case, rules) for y generations.

  • Parallel evolutionary modelling of geological processes

    2007, Parallel Computing
    Citation Excerpt :

    The complete list of the GA specifications and parameters is reported in Table 7. Note that, even if elitistic GAs can generally drive the search towards a sub-optimal solution (as they guarantee the survival of the fittest individuals – which may represent local optima – over the GA steps), they are commonly adopted for CA calibration purposes (e.g. cf. [31]). Moreover, elitistic GAs demonstrated to be a good solution for Di Gregorio based CA models.

  • The HPP rule with memory and the density classification task

    2010, International Journal of Modern Physics C
  • A very effective density classifier two-dimensional cellular automaton with memory

    2009, Journal of Physics A: Mathematical and Theoretical
View all citing articles on Scopus
View full text