David B.

Fogel Historic
Natural Selection Inc., USA Perspective

Nils Barricelli—Artificial Life, Coevolution, Self-Adaptation

ils Aall Barricelli was one of the 3. Reproduction may occur more

N
duce without requiring other organ-
pioneers in evolutionary compu- than once if x i+m,g = r (where isms of a different pattern, which he
tation. His publication of r
= 0, which designated an empty described using the term “another
“Esempi Numerici di processi di square), then x i+ r,g+ l = n , again species.” Conversely, if an organism
evoluzione,” a study in what would with the exception of a collision. was unable to reproduce without a
now be called artificial life, in the journal 4. If two numeric elements collide in continuous supply of other numbers
Methodos, in 1954 [1] is perhaps the ear- a cell, then if they are equal, only then it was dependent, and described as
liest published record of an evolution- one copy of the number is placed a “parasite.” Barricelli noted patterns of
ary simulation. The paper was in the cell. If the numbers are not recombination, including a multiple-
republished in 1957 in English [2], and equal, other rules are applied to point operation in which two patterns
detailed the results of programs that determine which number to place would collide and the result would be
were run at the Institute for Advanced in the cell or other cells. a new self-sustaining pattern with
Study in Princeton, NJ, in 1953.
Barricelli’s initial experiments com-
prised a simple simulation of numbers in Barricelli’s search for emergent patterns is reminiscent of the
a grid. The numbers moved in the grid
according to local rules that were speci- search for emergent properties in complex adaptive systems that
fied for each number. For example, pervaded artificial life research in the late 1980s and early 1990s.
suppose that a one-dimensional grid had
N = 20 cells and numbers were distrib-
uted in those cells at the initial genera- Figure 1, taken from [2], shows an numbers chosen from each of the “par-
tion g = 0. For g = 1, the numbers example of numbers propagating in ents.” A two-dimensional version of
would shift to different cells based on time through a series of 20-cell grids. the experiment was also performed,
the arrangement of numbers at g = 0, The numbers involved are –3, 1, and 5, with obvious similarities to the work of
and then progress further or “migrate” and also 0 representing an empty square. Conway’s Game of Life, popularized
based on their then-current positions. Starting from the arrangement at much later in [4]. Overall, Barricelli’s
To explain Barricelli’s specific g = 0 at the top of the grid (usually search for emergent patterns is reminis-
migration rules, say x i,g is the numeric found by using a set of playing cards), cent of the search for emergent proper-
entry at the gth generation in cell i. by the fourth generation, the pattern (5, ties in complex adaptive systems that
Barricelli [2] used rules such as: –3, 1, –3, 0, –3, 1) appears and persists pervaded artificial life research in the
1. A number shifts n cells to the right in every other generation. This example late 1980s and early 1990s.
if it is positive, or |n| cells to the is a “flat” grid, but Barricelli [2] noted In the late 1950s and early 1960s,
left if it is negative, unless this experiments with 512 cells in a tubular Barricelli extended his study beyond
results in a “collision” with another design (connecting the left and right emergent properties in deterministic
number (in which two numbers edges). simulations to playing games between
arrive at the same cell location). Barricelli made several observations strategies [5], [6], particularly with the
2. The same number x i,g = n may about the patterns that emerged from idea of determining the outcome of
reproduce m cells to the right (or such simple rules, which he termed “collisions” when numbers entered cells
left) if x i+n,g = m, again with the “organisms.” Organisms were defined in a grid simultaneously. Barricelli stud-
exception of a collision. to be independent if they could repro- ied the game Tac Tix, which is much

1556-603X/06/$20.00©2006IEEE FEBRUARY 2006 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 41

like Nim. The game is played in a 6 × 6 is a truly remarkable contribution to the Furthermore, Reed et al. [7] tested
square of 36 coins in which players history of evolutionary computation. both one-point crossover and uniform
alternate in removing one or more coins The framework studied in [7] crossover (22 years before uniform
in a single row or column. Only an involved a simple two-player card game crossover was reintroduced in [8]). A
uninterrupted series of coins can be with cards having two different values coevolving population of up to 10,000
removed. The player who removes the (low or high) and three possible bets individuals was used in multiple trials.
last coin loses. Using an IBM 704, (pass, low, or high, which cost 2, 3, or 7 The use of such a large population was
Barricelli represented the game state as a units, respectively). The players had to another notable facet of this work.
binary number and evolved a so-called decide which bet to make with which The game was sufficiently simple to
“decision function” that determined the card. If the players bet differently, the allow calculating the optimal strategy for
next move for a player. player that bet more won. If the players each situation using Von Neumann’s
game theory. The evolved low-hand
strategy was “usually close to the opti-
3 1 3
mum values” [7]. In one case, 90 per-
5 5 1
cent of the population had evolved
5 1 3 x 3 1 x 3 3
optimal strategies.
5 1 3 3 1 5 3 1 3 3 Barricelli’s contribution in [7], much
3 3 1 5 3 1 3 3 1 5 3 3 like the independent contribution of
5 3 1 3 3 1 5 3 1 3 3 5
[11], was one of the first demonstrable
successes of coevolution. Most
3 1 3 3 1 5 3 1 3 3 1 5 3 1
researchers are more familiar with the
3 1 5 3 1 3 3 1 5 3 1 3 3 work of [12] and in particular [13] as
5 3 1 3 3 1 5 3 1 3 3 1 5 3 pioneering efforts in coevolution.
1 3 3 1 5 3 1 3 3 1 5 3 1 3 3
Indeed these works were pioneering,
but consider the pioneering nature of
3 1 5 3 1 3 3 1 5 3 1 3 3 1 5
applying a population of 10,000 strate-
5 3 1 3 3 1 5 3 1 3 3 1 5 3 1 3 gies in a coevolutionary setting, using
3 3 1 5 3 1 3 3 1 5 3 1 3 3 1 self-adaptation, different forms of
recombination and mutation, on a com-
puter that is about one-half the speed of
FIGURE 1 A series of generations of numbers moving in a grid following the rules of [2]. The
first generation is shown in the top row of cells. Subsequent generations are shown as an Apple II, and doing this about
successive descending rows of cells. By the fourth generation, the pattern (5, –3, 1, –3, 0, 40 years ago. Even in the Tac Tix game
–3, 1) appears and persists in each subsequent generation. Note that a number with an [6], 10,000 generations of evolutionary
underscore has a negative value. The figure is adapted from [2].
required 80 hours. These were some
truly visionary experiments.
Barricelli measured the evolutionary bet the same level, then the player with When preparing Evolutionary Com-
progress of his coevolutionary gaming the higher-valued card won. Otherwise, putation: The Fossil Record [3], I was
simulation and concluded that the the result was a tie. fortunate to locate and speak with
resulting strategies had improved but The representation for playing Robert Toombs (co-author of [7]). I
were not as effective as human begin- strategies included eight positive num- asked specifically about the insight
ners. In spite of the obvious chasm, Bar- bers: four of these affected the player’s into using self-adaptation, since the
ricelli [6] speculated on the potential of genetic search properties and the other effort in [7] was, along with [9], per-
this coevolutionary method to address four defined the betting probabilities haps the earliest published work in
chess, writing “the only way to decide under different outcomes. The four this direction. Toombs [10] indicated
the question is to try.” It would require parameters that affected the player’s that the idea to use self-adaptation
another 40 years before some answers search included determining mutation grew from Barricelli’s earlier work on
would be available [14]–[17]. probabilities, a step size, and the proba- emergent patterns from migrating
Later, in the mid-1960s, Barricelli bility that a player would undergo numbers in a grid. Different mutation
collaborated with Jon Reed of Oslo crossover with another strategy. That is, rules were observed to give rise to
University and Robert Toombs, a the representation included not only different patterns, leading to the sug-
student in the Department of Genetics the encoding for how to bet, but also gestion that the computer could be
at the University of Washington how to search for new strategies, used to evolve appropriate variation
(Barricelli’s home location) in a simula- including the possibility for recombin- rules at the same time as searching for
tion that involved not only coevolution, ing with other strategies. This is now optimal game strategies.
but also self-adaptation. Reed et al. [7] known commonly as “self-adaptation.” (continued on page 45)

42 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | FEBRUARY 2006

Developmental Tools (continued from page 38)
to the main list (using Web-based tem- as long as this is possible and as long as it tionary-based multiobjective optimization techniques,”
Knowledge and Information Systems. An International Jour-
plates). It would also be very useful to is useful for people interested in evolu- nal, vol. 1, no. 3, pp. 269–308, Aug. 1999.
develop a more detailed database that tionary multi-objective optimization. [2] C.A. Coello Coello and N.C. Cortés, “Solving multiob-
jective optimization problems using an artificial immune
allows more precise statistics from the Needless to say, any contributions system,” Genetic Programming and Evolvable Machines,
papers in the repository. It is worth not- to the EMOO repository as well as vol. 6, no. 2, pp. 163–190, June 2005.
[3] C.A. Coello Coello and G.T. Pulido, “Multiobjective
ing that the updates of the main list of suggestions regarding how to improve Optimization using a Micro-Genetic Algorithm,” In Lee Spec-
references is still done by hand (which it, are welcome. tor et al., editor, Proceedings of the Genetic and Evolutionary
Computation Conference (GECCO’2001), pp. 274–282, San
explains the delays that sometimes arise). Francisco, California, Morgan Kaufmann Publishers, 2001.
Nevertheless, the many positive e-mail Acknowledgments [4] C.A. Coello Coello, D.A. Van Veldhuizen, and G.B.
Lamont, Evolutionary Algorithms for Solving Multi-
messages and oral communications that I The author acknowledges support from Objective Problems, Kluwer Academic Publishers, New
have received regarding the usefulness of CONACyT through project 42435-Y. York, May 2002. ISBN 0-3064-6762-3.
the EMOO repository motivates me to [5] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A
fast and elitist multiobjective genetic algorithm: NSGA-II,”
continue updating it. Therefore, I will References IEEE Transactions on Evolutionary Computation, vol. 6,
[1] C.A. Coello Coello, “A comprehensive survey of evolu- no. 2, pp. 182–197, Apr. 2002.
continue doing any updates required for

Historic Perspective (continued from page 42)

Careful readers will note that the research which occurred at the dawn of self-reproducing numeric patterns by data processing
machines, effects of hereditary control, mutation type and
title to [7] indicates that it is “part I.” modern computing. crossing,” J. Theoret. Biol., vol. 17, pp. 319–342, 1967.
Toombs told me [10] that no second I regret not having the opportunity [8] G. Syswerda, “Uniform crossover in genetic
algorithms,” Proc. 3rd Intern. Conf. on Genetic Algo-
part was ever published, which is a pity to spend some time with him personally, rithms, J.D. Schaffer (ed.), Morgan Kaufmann, San Mateo,
as there remains much to do in carry- but we can all spend some time with CA, pp. 2–9, 1989.
[9] R.S. Rosenberg, “Simulation of genetic populations
ing this research forward. For example, him impersonally by reading his publica- with biochemical properties,” Ph.D. Dissertation, Univ.
what would happen with alternative tions and extending his work. Michigan, Ann Arbor, MI, 1967.
[10] R. Toombs, personal communication, 1997.
representations, different variation [11] L.J. Fogel and G.H. Burgin, “Competitive goal-seeking
operators, and more complex card Acknowledgments through evolutionary programming,” Final report under
contract AF, vol. 19 no. 628, p. 5927, Air Force Cambridge
games? What types of games would be Portions of this article were drawn from Research Labs, 1969.
conducive to solutions by one-point [3], in accordance with IEEE copy- [12] R. Axelrod, “The evolution of strategies in the iterated
prisoner’s dilemma,” Genetic Algorithms and Simulated
crossover or uniform crossover, or rights. A more detailed discussion of [7] Annealing, L. Davis (ed.), Pitman, London, pp. 32–41, 1987.
none at all? (Reed et al. [7] remark on can be found in [18]. [13] W.D. Hillis, “Co-evolving parasites improve simulated
evolution as an optimization procedure,” Emergent Compu-
this question directly.) These are the tation, S. Forrest (ed.), MIT Press, Cambridge, MA, pp.
types of questions that can be answered 228–234, 1991.
References [14] G. Kendall and G. Whitwell, “An evolutionary
rapidly with modern computing. Moti- [1] N.A. Barricelli, “Esempi Numerici di processi di approach for tuning of a chess evaluation function using
evoluzione,” Methodos, pp. 45–68, 1954. population dynamics,” Proc. 2001 Congress on Evolution-
vated readers will find [5]–[7] reprinted [2] N.A. Barricell, “Symbiogenetic evolution processes ary Computation, IEEE Press, NY, pp. 995–1002, 2001.
in their entirety in [3]. realized by artificial methods,” Methodos, vol. 9, no. 35–36, [15] A. Hauptman and M. Sipper, “GP-EndChess: Using
pp. 143–182, 1957. genetic programming to evolve chess endgame players,”
Nils Barricelli remains a virtually [3] D.B. Fogel (Ed.), Evolutionary Computation: The Fossil Proc. 8th Europ. Conf. Genetic Programming, M. Keijzer
unknown figure in the evolutionary Record, IEEE Press, Piscatway, NJ, 1998. et al. (eds.), Lecture Notes in Computer Science, Springer,
[4] A.K. Dewdney, “Computer recreations: The game of life Berlin, vol. 3447, pp. 120–131, 2005.
computation community, but that is no acquires some successors in three dimensions,” Scientific [16] D.B. Fogel, T.J. Hays, S.L. Hahn, and J. Quon, “A self-
reflection on his pioneering contribu- American, vol. 256, no. 2, pp. 16–24. learning evolutionary chess program,” Proc. IEEE, vol. 92,
[5] N.A. Barricelli, “Numerical testing of evolution theories: no. 12, pp. 1947–1954, 2004.
tions and creative thoughts. He passed Part I: Theoretical introduction and basic tests,” Acta Bio- [17] D.B. Fogel, T.J. Hays, S.L. Hahn, and J. Quon, “Fur-
away on January 27, 1993, at the age of theoretica, vol. 16, no. 1–2, pp. 69–98, 1962. ther evolution of a self-learning chess program,” Proc. 2005
[6] N.A. Barricelli, “Numerical testing of evolution theories: IEEE Symp. Computational Intelligence and Games, G.
81, several years after the resurgence of Part II: Preliminary tests of performance, symbiogenesis and Kendall, and S. Lucas (chairs), IEEE Press, NY, 2005,
interest in artificial life, but with barely terrestrial life,” Acta Biotheoretica, vol. 16, no. 3–4, pp. 73–77, 2005.
pp. 99–126, 1963. [18] D.B. Fogel, “Unearthing a fossil from the history of
any recognition from that resurgence, [7] J. Reed, R. Toombs, and N.A. Barricelli, “Simulation of evolutionary computation,” Fundamenta Informaticae, vol.
and even fewer real references to his biological evolution and machine learning: I. Selection of 35 no. 1–2, pp. 1–16, 1998.

FEBRUARY 2006 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 45